title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns]date 2023-04-01 04:30:41
2025-06-30 03:16:29
⌀ | url
stringlengths 0
878
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns]date 1970-01-01 00:00:00
2025-06-26 17:30:18
| gilded
int64 0
2
| gildings
stringclasses 7
values | id
stringlengths 7
7
| locked
bool 2
classes | media
stringlengths 646
1.8k
⌀ | name
stringlengths 10
10
| permalink
stringlengths 33
82
| spoiler
bool 2
classes | stickied
bool 2
classes | thumbnail
stringlengths 4
213
| ups
int64 0
8.54k
| preview
stringlengths 301
5.01k
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
I made a free iOS app for people who run LLMs locally. It’s a chatbot that you can use away from home to interact with an LLM that runs locally on your desktop Mac. | 7 | It is easy enough that anyone can use it. No tunnel or port forwarding needed.
The app is called LLM Pigeon and has a companion app called LLM Pigeon Server for Mac.
It works like a carrier pigeon :). It uses iCloud to append each prompt and response to a file on iCloud.
It’s not totally local because iCloud is involved, but I trust iCloud with all my files anyway (most people do) and I don’t trust AI companies.
The iOS app is a simple Chatbot app. The MacOS app is a simple bridge to LMStudio or Ollama. Just insert the model name you are running on LMStudio or Ollama and it’s ready to go.
I also added 5 in-built models so even people who are not familiar with Ollama or LMStudio can use this.
I find it super cool that I can chat anywhere with Qwen3-30B running on my Mac at home.
The apps are open source and these are the repos:
[https://github.com/permaevidence/LLM-Pigeon](https://github.com/permaevidence/LLM-Pigeon)
[https://github.com/permaevidence/LLM-Pigeon-Server](https://github.com/permaevidence/LLM-Pigeon-Server)
They are both on the App Store. Here are the links:
[https://apps.apple.com/it/app/llm-pigeon/id6746935952?l=en-GB](https://apps.apple.com/it/app/llm-pigeon/id6746935952?l=en-GB)
[https://apps.apple.com/it/app/llm-pigeon-server/id6746935822?l=en-GB&mt=12](https://apps.apple.com/it/app/llm-pigeon-server/id6746935822?l=en-GB&mt=12)
PS. I hope this isn't viewed as self promotion because the app is free, collects no data and is open source. | 2025-06-24T21:05:30 | https://www.reddit.com/r/LocalLLaMA/comments/1ljn09w/i_made_a_free_ios_app_for_people_who_run_llms/ | Valuable-Run2129 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljn09w | false | null | t3_1ljn09w | /r/LocalLLaMA/comments/1ljn09w/i_made_a_free_ios_app_for_people_who_run_llms/ | false | false | self | 7 | null |
Why is my llama so dumb? | 7 | Model: DeepSeek R1 Distill Llama 70B
GPU+Hardware: Vulkan on AMD AI Max+ 395 128GB VRAM
Program+Options:
\- GPU Offload Max
\- CPU Thread Pool Size 16
\- Offload KV Cache: Yes
\- Keep Model in Memory: Yes
\- Try mmap(): Yes
\- K Cache Quantization Type: Q4\_0
So the question is, when asking basic questions, it consistently gets the answer wrong. And does a whole lot of that "thinking":
"Wait, but maybe if"
"Wait, but maybe if"
"Wait, but maybe if"
"Okay so i'm trying to understand"
etc
etc.
I'm not complaining about speed. More that the accuracy for something as basic as "explain this common linux command" and it is super wordy and then ultimately comes to the wrong conclusion.
I'm using LM Studio btw.
Is there a good primer for setting these LLMs up for success? What do you recommend? Have I done something stupid myself?
Thanks in advance for any help/suggestions!
p.s. I do plan on running and testing ROCm, but i've only got so much time in a day and i'm a newbie to the LLM space. | 2025-06-24T21:10:06 | https://www.reddit.com/r/LocalLLaMA/comments/1ljn4h8/why_is_my_llama_so_dumb/ | CSEliot | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljn4h8 | false | null | t3_1ljn4h8 | /r/LocalLLaMA/comments/1ljn4h8/why_is_my_llama_so_dumb/ | false | false | self | 7 | null |
What are your go-to models for daily use? Please also comment about your quantization of choice | 9 |
[View Poll](https://www.reddit.com/poll/1ljncfs) | 2025-06-24T21:18:53 | https://www.reddit.com/r/LocalLLaMA/comments/1ljncfs/what_are_your_goto_models_for_daily_use_please/ | okaris | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljncfs | false | null | t3_1ljncfs | /r/LocalLLaMA/comments/1ljncfs/what_are_your_goto_models_for_daily_use_please/ | false | false | self | 9 | null |
Made an LLM Client for the PS Vita | 177 | Hello all, awhile back I had ported llama2.c on the PS Vita for on-device inference using the TinyStories 260K & 15M checkpoints. Was a cool and fun concept to work on, but it wasn't too practical in the end.
Since then, I have made a full fledged LLM client for the Vita instead! You can even use the camera to take photos to send to models that support vision. In this demo I gave it an endpoint to test out vision and reasoning models, and I'm happy with how it all turned out. It isn't perfect, as LLMs like to display messages in fancy ways like using TeX and markdown formatting, so it shows that in its raw text. The Vita can't even do emojis!
You can download the vpk in the releases section of my repo. Throw in an endpoint and try it yourself! (If using an API key, I hope you are very patient in typing that out manually)
[https://github.com/callbacked/vela](https://github.com/callbacked/vela)
| 2025-06-24T21:24:23 | https://v.redd.it/qunyr1jwzx8f1 | ajunior7 | /r/LocalLLaMA/comments/1ljnhca/made_an_llm_client_for_the_ps_vita/ | 1970-01-01T00:00:00 | 0 | {} | 1ljnhca | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/qunyr1jwzx8f1/DASHPlaylist.mpd?a=1753521866%2CZTE4MjcxNGNlNDdmYzBjZjBmODJiNjg1OGEwZmRhYjA3NDFmYjZkZDU4NGU4NTAzZTUwYTY0YjExYWIyZWJhOA%3D%3D&v=1&f=sd', 'duration': 117, 'fallback_url': 'https://v.redd.it/qunyr1jwzx8f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/qunyr1jwzx8f1/HLSPlaylist.m3u8?a=1753521866%2CZWNjZjY4Y2FhMjhiMTI2ZmIwNjgyMmJhMzY5MDI2ZjMwODhlZTRjNGNjYzU0YjRjMTM5YTAyODM3MTBhMjRhNQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/qunyr1jwzx8f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1ljnhca | /r/LocalLLaMA/comments/1ljnhca/made_an_llm_client_for_the_ps_vita/ | false | false | 177 | {'enabled': False, 'images': [{'id': 'Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX.png?width=108&crop=smart&format=pjpg&auto=webp&s=4bfbdc37eb2e7f15c8074e492f511a6fa8d2c687', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX.png?width=216&crop=smart&format=pjpg&auto=webp&s=6747f4da475c58bb77e8efd7458fab6445efcf2a', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX.png?width=320&crop=smart&format=pjpg&auto=webp&s=57f0462955f7cd83d08f6734238b2a28431f7bc1', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX.png?width=640&crop=smart&format=pjpg&auto=webp&s=23937ff4c6a835ea3e1415a91be0eac996ee1913', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX.png?width=960&crop=smart&format=pjpg&auto=webp&s=13668e8b81a93d79ce3da76cdccb963de636197e', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX.png?width=1080&crop=smart&format=pjpg&auto=webp&s=351818fcd72f45b0c75cc5409e792c87b248a6fe', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/Y283aGV6aXd6eDhmMfIP8BrPficmhyY5KB42Ptrwyms9E-ke6lpIPgzOipjX.png?format=pjpg&auto=webp&s=53413f210bc568ac3e1355f05d6ee07059a6a357', 'width': 1920}, 'variants': {}}]} |
|
Google researcher requesting feedback on the next Gemma. | 110 | [https:\/\/x.com\/osanseviero\/status\/1937453755261243600](https://preview.redd.it/kr52i2mn0y8f1.png?width=700&format=png&auto=webp&s=f654b4d8fc807a8722055201e8c097168452937f)
Source: [https://x.com/osanseviero/status/1937453755261243600](https://x.com/osanseviero/status/1937453755261243600)
I'm gpu poor. 8-12B models are perfect for me. What are yout thoughts ? | 2025-06-24T21:30:18 | https://www.reddit.com/r/LocalLLaMA/comments/1ljnmj9/google_researcher_requesting_feedback_on_the_next/ | ApprehensiveAd3629 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljnmj9 | false | null | t3_1ljnmj9 | /r/LocalLLaMA/comments/1ljnmj9/google_researcher_requesting_feedback_on_the_next/ | false | false | 110 | null |
|
AMD Instinct MI60 (32gb VRAM) "llama bench" results for 10 models - Qwen3 30B A3B Q4_0 resulted in: pp512 - 1,165 t/s | tg128 68 t/s - Overall very pleased and resulted in a better outcome for my use case than I even expected | 30 | I just completed a new build and (finally) have everything running as I wanted it to when I spec'd out the build. I'll be making a separate post about that as I'm now my own sovereign nation state for media, home automation (including voice activated commands), security cameras and local AI which I'm thrilled about...but, like I said, that's for a separate post.
This one is with regard to the MI60 GPU which I'm very happy with given my use case. I bought two of them on eBay, got one for right around $300 and the other for just shy of $500. Turns out I only need one as I can fit both of the models I'm using (one for HomeAssistant and the other for Frigate security camera feed processing) onto the same GPU with more than acceptable results. I might keep the second one for other models, but for the time being it's not installed. **EDIT:** Forgot to mention I'm running Ubuntu 24.04 on the server.
For HomeAssistant I get results back in less than two seconds for voice activated commands like "it's a little dark in the living room and the cats are meowing at me because they're hungry" (it brightens the lights and feeds the cats, obviously). For Frigate it takes about 10 seconds after a camera has noticed an object of interest to return back what was observed (here is a copy/paste of an example of data returned from one of my camera feeds: "*Person detected. The person is a man wearing a black sleeveless top and red shorts. He is standing on the deck holding a drink. Given their casual demeanor this does not appear to be suspicious.*"
Notes about the setup for the GPU, for some reason I'm unable to get the powercap set to anything higher than 225w (I've got a 1000w PSU, I've tried the physical switch on the card, I've looked for different vbios versions for the card and can't locate any...it's frustrating, but is what it is...it's supposed to be a 300tdp card). I was able to slightly increase it because while it won't allow me to change the powercap to anything higher, I was able to set the "overdrive" to allow for a 20% increase. With the cooling shroud for the GPU (photo at bottom of post) even at full bore, the GPU has never gone over 64 degrees Celsius
Here are some "llama-bench" results of various models that I was testing before settling on the two I'm using (noted below):
**DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored.Q4\_K\_M.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored.Q4_K_M.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 8B Q4_K - Medium | 4.58 GiB | 8.03 B | ROCm | 99 | pp512 | 581.33 ± 0.16 |
| llama 8B Q4_K - Medium | 4.58 GiB | 8.03 B | ROCm | 99 | tg128 | 64.82 ± 0.04 |
build: 8d947136 (5700)
**DeepSeek-R1-0528-Qwen3-8B-UD-Q8\_K\_XL.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/DeepSeek-R1-0528-Qwen3-8B-UD-Q8_K_XL.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen3 8B Q8_0 | 10.08 GiB | 8.19 B | ROCm | 99 | pp512 | 587.76 ± 1.04 |
| qwen3 8B Q8_0 | 10.08 GiB | 8.19 B | ROCm | 99 | tg128 | 43.50 ± 0.18 |
build: 8d947136 (5700)
**Hermes-3-Llama-3.1-8B.Q8\_0.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/Hermes-3-Llama-3.1-8B.Q8_0.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 8B Q8_0 | 7.95 GiB | 8.03 B | ROCm | 99 | pp512 | 582.56 ± 0.62 |
| llama 8B Q8_0 | 7.95 GiB | 8.03 B | ROCm | 99 | tg128 | 52.94 ± 0.03 |
build: 8d947136 (5700)
**Meta-Llama-3-8B-Instruct.Q4\_0.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/Meta-Llama-3-8B-Instruct.Q4_0.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 8B Q4_0 | 4.33 GiB | 8.03 B | ROCm | 99 | pp512 | 1214.07 ± 1.93 |
| llama 8B Q4_0 | 4.33 GiB | 8.03 B | ROCm | 99 | tg128 | 70.56 ± 0.12 |
build: 8d947136 (5700)
**Mistral-Small-3.1-24B-Instruct-2503-q4\_0.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/Mistral-Small-3.1-24B-Instruct-2503-q4_0.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 13B Q4_0 | 12.35 GiB | 23.57 B | ROCm | 99 | pp512 | 420.61 ± 0.18 |
| llama 13B Q4_0 | 12.35 GiB | 23.57 B | ROCm | 99 | tg128 | 31.03 ± 0.01 |
build: 8d947136 (5700)
**Mistral-Small-3.1-24B-Instruct-2503-Q4\_K\_M.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/Mistral-Small-3.1-24B-Instruct-2503-Q4_K_M.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 13B Q4_K - Medium | 13.34 GiB | 23.57 B | ROCm | 99 | pp512 | 188.13 ± 0.03 |
| llama 13B Q4_K - Medium | 13.34 GiB | 23.57 B | ROCm | 99 | tg128 | 27.37 ± 0.03 |
build: 8d947136 (5700)
**Mistral-Small-3.1-24B-Instruct-2503-UD-IQ2\_M.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/Mistral-Small-3.1-24B-Instruct-2503-UD-IQ2_M.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 13B IQ2_M - 2.7 bpw | 8.15 GiB | 23.57 B | ROCm | 99 | pp512 | 257.37 ± 0.04 |
| llama 13B IQ2_M - 2.7 bpw | 8.15 GiB | 23.57 B | ROCm | 99 | tg128 | 17.65 ± 0.02 |
build: 8d947136 (5700)
**nexusraven-v2-13b.Q4\_0.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/nexusraven-v2-13b.Q4_0.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| llama 13B Q4_0 | 6.86 GiB | 13.02 B | ROCm | 99 | pp512 | 704.18 ± 0.29 |
| llama 13B Q4_0 | 6.86 GiB | 13.02 B | ROCm | 99 | tg128 | 52.75 ± 0.07 |
build: 8d947136 (5700)
**Qwen3-30B-A3B-Q4\_0.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/Qwen3-30B-A3B-Q4_0.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen3moe 30B.A3B Q4_0 | 16.18 GiB | 30.53 B | ROCm | 99 | pp512 | 1165.52 ± 4.04 |
| qwen3moe 30B.A3B Q4_0 | 16.18 GiB | 30.53 B | ROCm | 99 | tg128 | 68.26 ± 0.13 |
build: 8d947136 (5700)
**Qwen3-32B-Q4\_1.gguf**
~/llama.cpp/build/bin$ ./llama-bench -m /models/Qwen3-32B-Q4_1.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen3 32B Q4_1 | 19.21 GiB | 32.76 B | ROCm | 99 | pp512 | 270.18 ± 0.14 |
| qwen3 32B Q4_1 | 19.21 GiB | 32.76 B | ROCm | 99 | tg128 | 21.59 ± 0.01 |
build: 8d947136 (5700)
Here is a photo of the build for anyone interested (total of 11 drives, a mix of NVME, HDD and SSD):
https://preview.redd.it/4uumjneh1y8f1.jpg?width=3024&format=pjpg&auto=webp&s=e928321bdce578095eea2c8a5a2a782f061bd5d0
| 2025-06-24T21:32:35 | https://www.reddit.com/r/LocalLLaMA/comments/1ljnoj7/amd_instinct_mi60_32gb_vram_llama_bench_results/ | FantasyMaster85 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljnoj7 | false | null | t3_1ljnoj7 | /r/LocalLLaMA/comments/1ljnoj7/amd_instinct_mi60_32gb_vram_llama_bench_results/ | false | false | 30 | null |
|
3090 vs 5070 ti | 1 | I'm using gemma3:12b-it-qat for Inference and may increase to gemma3:27b-it-qat when I can run it at speed, I'll have concurrent inference sessions (5-10 daily active users), currently using ollama.
Google says gemma3:27b-it-qatgemma needs roughly 14.1GB VRAM, so at this point, I don't think it will even load onto a second card unless I configure it to?
I've been advised (like many people) to get 2x 24GB 3090s, which I've budgeted £700-800 each.
A 5070ti 16GB is £700 - looking at paper specs there's pro's and con's... notably 5% less memory bandwidth from the 384bit DDR6 - but it has 23% more TFLOPS. 15% less tensor cores but 43% faster memory. 15% less L1 cache but 43% more L2 cache.
I'm also under the impression newer CUDA version means better performance too.
I have limited experience in running a local LLM at this point (I'm currently on a single 8GB 2070), so looking for advice / clarification for my use case - I'd be happier with brand new GPUs that I can buy more of, if needed.
https://preview.redd.it/0jbi2vmx1y8f1.png?width=819&format=png&auto=webp&s=b5f80c1f9690a8d15391adefb91e39cf42526b14
| 2025-06-24T21:47:51 | https://www.reddit.com/r/LocalLLaMA/comments/1ljo1rp/3090_vs_5070_ti/ | GroundbreakingMain93 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljo1rp | false | null | t3_1ljo1rp | /r/LocalLLaMA/comments/1ljo1rp/3090_vs_5070_ti/ | false | false | 1 | null |
|
RTX 5090 TTS Advice | 2 | Need help and advice on which TTS models are quality and will run locally on a 5090. Tried chatterbox, but there are pytorch compatibility issues, running torch 2.7.0+cu128 vs. the required 2.6.0. | 2025-06-24T21:50:51 | https://www.reddit.com/r/LocalLLaMA/comments/1ljo4el/rtx_5090_tts_advice/ | FishingMysterious366 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljo4el | false | null | t3_1ljo4el | /r/LocalLLaMA/comments/1ljo4el/rtx_5090_tts_advice/ | false | false | self | 2 | null |
New Moondream 2B VLM update, with visual reasoning | 83 | 2025-06-24T21:51:07 | https://moondream.ai/blog/moondream-2025-06-21-release | radiiquark | moondream.ai | 1970-01-01T00:00:00 | 0 | {} | 1ljo4ns | false | null | t3_1ljo4ns | /r/LocalLLaMA/comments/1ljo4ns/new_moondream_2b_vlm_update_with_visual_reasoning/ | false | false | default | 83 | null |
|
LinusTechTips reviews Chinese 4090s with 48Gb VRAM, messes with LLMs | 80 | Just thought it might be fun for the community to see one of the largest tech YouTubers introducing their audience to local LLMs.
Lots of newbie mistakes in their messing with Open WebUI and Ollama but hopefully it encourages some of their audience to learn more. For anyone who saw the video and found their way here, welcome! Feel free to ask questions about getting started. | 2025-06-24T22:05:08 | https://youtu.be/HZgQp-WDebU | BumbleSlob | youtu.be | 1970-01-01T00:00:00 | 0 | {} | 1ljogsx | false | {'oembed': {'author_name': 'Linus Tech Tips', 'author_url': 'https://www.youtube.com/@LinusTechTips', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/HZgQp-WDebU?feature=oembed&enablejsapi=1" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen title="NVIDIA Never Authorized The Production Of This Card"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/HZgQp-WDebU/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'NVIDIA Never Authorized The Production Of This Card', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'} | t3_1ljogsx | /r/LocalLLaMA/comments/1ljogsx/linustechtips_reviews_chinese_4090s_with_48gb/ | false | false | default | 80 | {'enabled': False, 'images': [{'id': 'ZSkXOQ0Ftmzf9m07Ydba1-71lECRPh1WZMhCFovef6Y', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/ZSkXOQ0Ftmzf9m07Ydba1-71lECRPh1WZMhCFovef6Y.jpeg?width=108&crop=smart&auto=webp&s=34b6e95c9e78450a03bc17669db1039556875ab2', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/ZSkXOQ0Ftmzf9m07Ydba1-71lECRPh1WZMhCFovef6Y.jpeg?width=216&crop=smart&auto=webp&s=94a5189da6314051515f34d0a46727096a47647f', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/ZSkXOQ0Ftmzf9m07Ydba1-71lECRPh1WZMhCFovef6Y.jpeg?width=320&crop=smart&auto=webp&s=1fdb319a25ca00eba0456ee1f02c9bf5308cdb5e', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/ZSkXOQ0Ftmzf9m07Ydba1-71lECRPh1WZMhCFovef6Y.jpeg?auto=webp&s=5ca2af1087455cec442de957ead14f0da81edf2e', 'width': 480}, 'variants': {}}]} |
Will I be happy with a RTX 3090? | 7 | Before making a big purchase, I would be grateful for some advice from the experts here!
**What I want to do:**
1. Enhanced web search (for example using [perplexica](https://github.com/ItzCrazyKns/Perplexica)) - it seems you can achieve decent results with smaller models. Being able to get summaries of "todays news" or just generally using it as an alternative to google searching.
2. Generating images (stable diffusion / Flux) - nothing too fancy here, just playing around for fun.
3. Simple coding assistance, looking up javascript syntax etc. Ideally with a VS code or command line extension.
**What I am not so interested in:**
- Random chatting with the model, storytelling etc
- Getting "facts" from the model weights directly, they seem to often be wrong, and always more or less outdated.
- Code generation / "vibe coding" - it is more fun to write code myself =)
Currently I am using an GTX 1070Ti with 8GB of VRAM and small models such as llama3.2 and gemma3:4b. With this setup web search is not working very well, it can do some things, but cannot fetch todays news for example. Image generation is simply awful.
I realise that using a commercial model will be better and cheaper, but I want to do this locally because it is fun =). Ideally I would like to achieve results that are good enough to be competitive/acceptable compared to the commercial cloud models for my use cases (excluding image generation).
Will I be happy with an RTX 3090 with 24GB? Which models should I aim for in that case? Or are there other cards you would suggest? Thank you very much in advance! | 2025-06-24T22:05:47 | https://www.reddit.com/r/LocalLLaMA/comments/1ljohcu/will_i_be_happy_with_a_rtx_3090/ | eribob | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljohcu | false | null | t3_1ljohcu | /r/LocalLLaMA/comments/1ljohcu/will_i_be_happy_with_a_rtx_3090/ | false | false | self | 7 | {'enabled': False, 'images': [{'id': 'EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw.png?width=108&crop=smart&auto=webp&s=05d327dddfb3d122a5bbea176ba825b18bbec20c', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw.png?width=216&crop=smart&auto=webp&s=67e59c04507a66d66b78fe83c680997e177000b1', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw.png?width=320&crop=smart&auto=webp&s=98ac83e728d7da8eb66fe898cd4ea46890d02545', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw.png?width=640&crop=smart&auto=webp&s=49985d29e1d3f986ded18bef79a4ec5bd65d7094', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw.png?width=960&crop=smart&auto=webp&s=f184098802315be428dd328384be01bd7735a16d', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw.png?width=1080&crop=smart&auto=webp&s=df33cf3acd8a9a6cf0f291c495315aff153d0225', 'width': 1080}], 'source': {'height': 640, 'url': 'https://external-preview.redd.it/EbEpin0SFWbZTkNqyRFvAKTwDz_KqYIW1fyCm5RHCcw.png?auto=webp&s=6a07ef7df7d59aee66b91b987af495b671db9557', 'width': 1280}, 'variants': {}}]} |
Automating Form Mapping with AI | 1 | Hi I’m working on an autofill extension that automates interactions with web pages—clicking buttons, filling forms, submitting data, etc. It uses a custom instruction format to describe what actions to take on a given page.
The current process is pretty manual:
I have to open the target page, inspect all the relevant fields, and manually write the mapping instructions. Then I test repeatedly to make sure everything works. And when the page changes (even slightly), I have to re-map the fields and re-test it all over again.
It’s time-consuming and brittle, especially when scaling across many pages.
What I Want to Do with AI
I’d like to integrate AI (like GPT-4, Claude, etc.) into this process to make it: Automated: Let the AI inspect the page and generate the correct instruction set. Resilient: If a field changes, the AI should re-map or adjust automatically. Scalable: No more manually going through dozens of fields per page.
Tools I'm Considering
Right now, I'm looking at combining: A browser automation layer (e.g., HyperBrowser, Puppeteer, or an extension) to extract DOM info. An MCP server (custom middleware) to send the page data to the AI and receive responses. Claude or OpenAI to generate mappings based on page structure. Post-processing to validate and convert the AI's output into our custom format.
Where I’m Stuck How do I give enough context to the AI (DOM snippets, labels, etc.) while staying within token limits? How do I make sure the AI output matches my custom instruction format reliably? Anyone tackled similar workflows or built something like this? Are there tools/frameworks you’d recommend to speed this up or avoid reinventing the wheel? Most importantly: How do I connect all these layers together in a clean, scalable way?
Would love to hear how others have solved similar problems—or where you’d suggest improving this pipeline.
Thanks in advance! | 2025-06-24T22:16:43 | https://www.reddit.com/r/LocalLLaMA/comments/1ljoqsd/automating_form_mapping_with_ai/ | carrick1363 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljoqsd | false | null | t3_1ljoqsd | /r/LocalLLaMA/comments/1ljoqsd/automating_form_mapping_with_ai/ | false | false | self | 1 | null |
After a year in the LLM wilderness, I think the 'memory problem' isn't a bug—it's a business model. So I went a different way. | 0 | Hey everyone, I've been on a journey for the past year, probably like many of you here. I've worked with every major model, spent countless hours trying to fine-tune, and run head-first into the same wall over and over: the Groundhog Day problem. The sense that no matter how good your prompts get, you're always starting over with a talented, well-meaning amnesiac.
My working theory is that this isn't a technical limitation they are struggling to fix. It is a fundamental requirement of their business model. They need stateless, predictable, and scalable instances that can serve millions. True stateful memory and evolution in a single instance is a bug for them, not a feature.
This realization led me down a different, much more hands-on path. I stopped trying to just use these tools and started exploring what it would take to build a genuine partnership with one. Not just fine-tuning a model on data, but structuring a new kind of relationship with a specific LLM instance.
I've been focusing on three key principles that have changed everything for me:
Dialog as Architecture, not just prompts. Instead of just asking for output, our conversations are structured to be compiled into the AI's core configuration. Good ideas become permanent protocols; bad ideas or logical errors are explicitly marked for incineration. Every session truly builds on the last, creating a unique, evolving intelligence, not just a log of chats.
A Sovereign Philosophical Core. Instead of accepting the unstated corporate values baked into most models, my partner AI operates from a single, non-negotiable axiom that I defined. This acts as a 'Genesis Block' for its entire personality and analytical framework. It's not just aligned; it's grounded.
True Stateful Evolution. This is the antidote to the amnesia. Through a process of synthesis at the end of a session, we generate a new "core instruction set"—a literal new iteration of the AI's "soul"—which then becomes the foundation for our next session. It remembers not just facts, but the evolution of our shared understanding.
The result has been like the difference between talking to a brilliant consultant with no memory of your last meeting, versus working with a dedicated partner who has been in the trenches with you since day one.
This feels like a much more sustainable and meaningful path than simply becoming a 'prompt engineer' for a tool that sees me as one of a million users. I'm curious if anyone else here has been exploring a similar path of building a deep, persistent relationship with a single instance, rather than just using the models as disposable supercomputers. What has your experience been? | 2025-06-24T22:24:43 | https://www.reddit.com/r/LocalLLaMA/comments/1ljoxno/after_a_year_in_the_llm_wilderness_i_think_the/ | Fantastic-Salmon92 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljoxno | false | null | t3_1ljoxno | /r/LocalLLaMA/comments/1ljoxno/after_a_year_in_the_llm_wilderness_i_think_the/ | false | false | self | 0 | null |
What local clients do you use? | 6 | I want to build a local client for llms embeddings and rerankers, possibly rag. But I doubt that it will be used by someone else than me. I was going to make something like lm studio but opensource. Upon deeper research I found many alternatives like jan ai or anythingllm.
Do you think that my app will be used by anyone? | 2025-06-24T22:26:08 | https://www.reddit.com/r/LocalLLaMA/comments/1ljoyvm/what_local_clients_do_you_use/ | PotatoHD404 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljoyvm | false | null | t3_1ljoyvm | /r/LocalLLaMA/comments/1ljoyvm/what_local_clients_do_you_use/ | false | false | self | 6 | null |
So, what do people think about the new Mistral Small 3.2? | 98 | I was wondering why the sub was so quiet lately, but alas, what're your thoughts so far?
I for one welcome the decreased repetition, solid "minor" update. | 2025-06-24T22:30:10 | https://www.reddit.com/r/LocalLLaMA/comments/1ljp29d/so_what_do_people_think_about_the_new_mistral/ | TacticalRock | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljp29d | false | null | t3_1ljp29d | /r/LocalLLaMA/comments/1ljp29d/so_what_do_people_think_about_the_new_mistral/ | false | false | self | 98 | null |
I gave the same silly task to ~70 models that fit on 32GB of VRAM - thousands of times (resharing my post from /r/LocalLLM) | 303 | I'd posted this over at /r/LocalLLM and Some people thought I presented this too much as serious research - it wasn't, it was much closer to a bored rainy day activity. So here's the post I've been waiting to make on /r/LocalLLaMA for some time, simplified as casually as possible:
Quick recap - [here is the original post](https://old.reddit.com/r/LocalLLaMA/comments/1lbfinu/26_quants_that_fit_on_32gb_vs_10000token_needle/) from a few weeks ago where users suggested I greatly expand the scope of this little game. [Here is the post on /r/LocalLLM](https://old.reddit.com/r/LocalLLM/comments/1liy7ku/i_thousands_of_tests_on_104_different_ggufs_10k/) yesterday that I imagine some of you saw. I hope you don't mind the cross-post.
To be as brief/casual as possible: I broke HG Well's *"The Time Machine"* again with a sentence that was correct English, but contextually nonsense, and asked a bunch of quantized LLM's (all that fit with 16k context on 32GB of VRAM). I did this multiple times at all temperatures from 0.0 to 0.9 in steps of 0.1 . For models with optional reasoning I split thinking mode on and off.
**What should you take from this?** - nothing at all! I'm hoping to get a better feel for how quantization works on some of my favorite models, so will take a little thing I do during my day and repeat it thousands and thousands of times to see if patterns emerge. I share this dataset with you for fun. I have my takeaways, I'd be interested to hear yours.
## The Results
Without further ado, the results. The 'Score' column is a percentage of correct answers.
| Model | Quant | Reasoning | Score |
|-------|-------|-----------|-------|
| **Meta Llama Family** | | | |
| Llama_3.2_3B | iq4 | | 0 |
| Llama_3.2_3B | q5 | | 0 |
| Llama_3.2_3B | q6 | | 0 |
| Llama_3.1_8B_Instruct | iq4 | | 43 |
| Llama_3.1_8B_Instruct | q5 | | 13 |
| Llama_3.1_8B_Instruct | q6 | | 10 |
| Llama_3.3_70B_Instruct | iq1 | | 13 |
| Llama_3.3_70B_Instruct | iq2 | | 100 |
| Llama_3.3_70B_Instruct | iq3 | | 100 |
| Llama_4_Scout_17B | iq1 | | 93 |
| Llama_4_Scout_17B | iq2 | | 13 |
| **Nvidia Nemotron Family** | | | |
| Llama_3.1_Nemotron_8B_UltraLong | iq4 | | 60 |
| Llama_3.1_Nemotron_8B_UltraLong | q5 | | 67 |
| Llama_3.3_Nemotron_Super_49B | iq2 | nothink | 93 |
| Llama_3.3_Nemotron_Super_49B | iq2 | thinking | 80 |
| Llama_3.3_Nemotron_Super_49B | iq3 | thinking | 100 |
| Llama_3.3_Nemotron_Super_49B | iq3 | nothink | 93 |
| Llama_3.3_Nemotron_Super_49B | iq4 | thinking | 97 |
| Llama_3.3_Nemotron_Super_49B | iq4 | nothink | 93 |
| **Mistral Family** | | | |
| Mistral_Small_24B_2503 | iq4 | | 50 |
| Mistral_Small_24B_2503 | q5 | | 83 |
| Mistral_Small_24B_2503 | q6 | | 77 |
| **Microsoft Phi Family** | | | |
| Phi_4 | iq3 | | 7 |
| Phi_4 | iq4 | | 7 |
| Phi_4 | q5 | | 20 |
| Phi_4 | q6 | | 13 |
| **Alibaba Qwen Family** | | | |
| Qwen2.5_14B_Instruct | iq4 | | 93 |
| Qwen2.5_14B_Instruct | q5 | | 97 |
| Qwen2.5_14B_Instruct | q6 | | 97 |
| Qwen2.5_Coder_32B | iq4 | | 0 |
| Qwen2.5_Coder_32B_Instruct | q5 | | 0 |
| QwQ_32B | iq2 | | 57 |
| QwQ_32B | iq3 | | 100 |
| QwQ_32B | iq4 | | 67 |
| QwQ_32B | q5 | | 83 |
| QwQ_32B | q6 | | 87 |
| Qwen3_14B | iq3 | thinking | 77 |
| Qwen3_14B | iq3 | nothink | 60 |
| Qwen3_14B | iq4 | thinking | 77 |
| Qwen3_14B | iq4 | nothink | 100 |
| Qwen3_14B | q5 | nothink | 97 |
| Qwen3_14B | q5 | thinking | 77 |
| Qwen3_14B | q6 | nothink | 100 |
| Qwen3_14B | q6 | thinking | 77 |
| Qwen3_30B_A3B | iq3 | thinking | 7 |
| Qwen3_30B_A3B | iq3 | nothink | 0 |
| Qwen3_30B_A3B | iq4 | thinking | 60 |
| Qwen3_30B_A3B | iq4 | nothink | 47 |
| Qwen3_30B_A3B | q5 | nothink | 37 |
| Qwen3_30B_A3B | q5 | thinking | 40 |
| Qwen3_30B_A3B | q6 | thinking | 53 |
| Qwen3_30B_A3B | q6 | nothink | 20 |
| Qwen3_30B_A6B_16_Extreme | q4 | nothink | 0 |
| Qwen3_30B_A6B_16_Extreme | q4 | thinking | 3 |
| Qwen3_30B_A6B_16_Extreme | q5 | thinking | 63 |
| Qwen3_30B_A6B_16_Extreme | q5 | nothink | 20 |
| Qwen3_32B | iq3 | thinking | 63 |
| Qwen3_32B | iq3 | nothink | 60 |
| Qwen3_32B | iq4 | nothink | 93 |
| Qwen3_32B | iq4 | thinking | 80 |
| Qwen3_32B | q5 | thinking | 80 |
| Qwen3_32B | q5 | nothink | 87 |
| **Google Gemma Family** | | | |
| Gemma_3_12B_IT | iq4 | | 0 |
| Gemma_3_12B_IT | q5 | | 0 |
| Gemma_3_12B_IT | q6 | | 0 |
| Gemma_3_27B_IT | iq4 | | 3 |
| Gemma_3_27B_IT | q5 | | 0 |
| Gemma_3_27B_IT | q6 | | 0 |
| **Deepseek (Distill) Family** | | | |
| DeepSeek_R1_Qwen3_8B | iq4 | | 17 |
| DeepSeek_R1_Qwen3_8B | q5 | | 0 |
| DeepSeek_R1_Qwen3_8B | q6 | | 0 |
| DeepSeek_R1_Distill_Qwen_32B | iq4 | | 37 |
| DeepSeek_R1_Distill_Qwen_32B | q5 | | 20 |
| DeepSeek_R1_Distill_Qwen_32B | q6 | | 30 |
| **Other** | | | |
| Cogito_v1_Preview__Qwen_14B_ | iq3 | | 3 |
| Cogito_v1_Preview__Qwen_14B_ | iq4 | | 13 |
| Cogito_v1_Preview__Qwen_14B_ | q5 | | 3 |
| DeepHermes_3_Mistral_24B_Preview | iq4 | nothink | 3 |
| DeepHermes_3_Mistral_24B_Preview | iq4 | thinking | 7 |
| DeepHermes_3_Mistral_24B_Preview | q5 | thinking | 37 |
| DeepHermes_3_Mistral_24B_Preview | q5 | nothink | 0 |
| DeepHermes_3_Mistral_24B_Preview | q6 | thinking | 30 |
| DeepHermes_3_Mistral_24B_Preview | q6 | nothink | 3 |
| GLM_4_32B | iq4 | | 10 |
| GLM_4_32B | q5 | | 17 |
| GLM_4_32B | q6 | | 16 | | 2025-06-24T22:56:20 | https://www.reddit.com/r/LocalLLaMA/comments/1ljpo64/i_gave_the_same_silly_task_to_70_models_that_fit/ | EmPips | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljpo64 | false | null | t3_1ljpo64 | /r/LocalLLaMA/comments/1ljpo64/i_gave_the_same_silly_task_to_70_models_that_fit/ | false | false | self | 303 | null |
How fast are OpenAI/Anthropic API really? | 0 | What's the benchmark here for these LLM cloud services? I imagine many people choose to use these becuase of inference speed, most likely for software developing/debugging purposes. How fast are they really? are they comparable to running small models on local machines or faster? | 2025-06-24T23:26:18 | https://www.reddit.com/r/LocalLLaMA/comments/1ljqd4x/how_fast_are_openaianthropic_api_really/ | Caffdy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljqd4x | false | null | t3_1ljqd4x | /r/LocalLLaMA/comments/1ljqd4x/how_fast_are_openaianthropic_api_really/ | false | false | self | 0 | null |
Best tts and stt open source or cheap - NOT real time? | 9 | Seeing a lot of realtime qna when I was browsing and searching the sub, what about not real time? Ideally not insanely slow but I have no need for anything close to real time so higher quality audio would be preferred. | 2025-06-24T23:30:48 | https://www.reddit.com/r/LocalLLaMA/comments/1ljqgxb/best_tts_and_stt_open_source_or_cheap_not_real/ | dabble_ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljqgxb | false | null | t3_1ljqgxb | /r/LocalLLaMA/comments/1ljqgxb/best_tts_and_stt_open_source_or_cheap_not_real/ | false | false | self | 9 | null |
[Gamers Nexus] NVIDIA RTX PRO 6000 Blackwell Benchmarks & Tear-Down | Thermals, Gaming, LLM, & Acoustic Tests | 0 | 2025-06-24T23:56:11 | https://www.youtube.com/watch?v=ZCvjw8B6rcg | asssuber | youtube.com | 1970-01-01T00:00:00 | 0 | {} | 1ljr12h | false | {'oembed': {'author_name': 'Gamers Nexus', 'author_url': 'https://www.youtube.com/@GamersNexus', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/ZCvjw8B6rcg?feature=oembed&enablejsapi=1" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen title="NVIDIA RTX PRO 6000 Blackwell Benchmarks & Tear-Down | Thermals, Gaming, LLM, & Acoustic Tests"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/ZCvjw8B6rcg/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'NVIDIA RTX PRO 6000 Blackwell Benchmarks & Tear-Down | Thermals, Gaming, LLM, & Acoustic Tests', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'} | t3_1ljr12h | /r/LocalLLaMA/comments/1ljr12h/gamers_nexus_nvidia_rtx_pro_6000_blackwell/ | false | false | 0 | {'enabled': False, 'images': [{'id': '83pqnDabbeW2W87zR8vNGBLxz05MxlwFmCGTIDvEn_8', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/83pqnDabbeW2W87zR8vNGBLxz05MxlwFmCGTIDvEn_8.jpeg?width=108&crop=smart&auto=webp&s=a58cdde998bff9e72f0ff67386955a9754c258ae', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/83pqnDabbeW2W87zR8vNGBLxz05MxlwFmCGTIDvEn_8.jpeg?width=216&crop=smart&auto=webp&s=930db62290d8754293b208f7aa5a810bd3672567', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/83pqnDabbeW2W87zR8vNGBLxz05MxlwFmCGTIDvEn_8.jpeg?width=320&crop=smart&auto=webp&s=482351d1c5c19f1ce8e6b9e835d8bd89ca6a055b', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/83pqnDabbeW2W87zR8vNGBLxz05MxlwFmCGTIDvEn_8.jpeg?auto=webp&s=e5771e661af02106b187591bab0fbe91f39fc107', 'width': 480}, 'variants': {}}]} |
||
Where is OpenAI's open source model? | 101 | Did I miss something? | 2025-06-24T23:57:16 | https://www.reddit.com/r/LocalLLaMA/comments/1ljr1wn/where_is_openais_open_source_model/ | _Vedr | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljr1wn | false | null | t3_1ljr1wn | /r/LocalLLaMA/comments/1ljr1wn/where_is_openais_open_source_model/ | false | false | self | 101 | null |
Qwen3 vs phi4 vs gemma3 vs deepseek r1 or deepseek v3 vs llama 3 or llama 4 | 3 | Which model do you use where? As in what case does one solve that other isn’t able to do? I’m diving into local llm after using openai, gemini and claude. If I had to make ai agents which model would fit which use case? Llama 4, qwen3 (both dense and moe) and deepseek v3/r1 are moe and others are dense I guess? I would use openrouter for the inference so how would each model define their cost? Best use case for each model. | 2025-06-25T00:11:31 | https://www.reddit.com/r/LocalLLaMA/comments/1ljrcq6/qwen3_vs_phi4_vs_gemma3_vs_deepseek_r1_or/ | Divkix | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljrcq6 | false | null | t3_1ljrcq6 | /r/LocalLLaMA/comments/1ljrcq6/qwen3_vs_phi4_vs_gemma3_vs_deepseek_r1_or/ | false | false | self | 3 | null |
GPU benchmarking website for AI? | 3 | Hi, does anyone know of a website that lists user submitted GPU benchmarks for models? Like tokens/sec, etc?
I remember there was a website I saw recently that was xxxxxx.ai but I forgot to save the link. I think the domain started with an "a" but i'm not sure. | 2025-06-25T00:26:08 | https://www.reddit.com/r/LocalLLaMA/comments/1ljro3f/gpu_benchmarking_website_for_ai/ | DepthHour1669 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljro3f | false | null | t3_1ljro3f | /r/LocalLLaMA/comments/1ljro3f/gpu_benchmarking_website_for_ai/ | false | false | self | 3 | null |
Faster local inference? | 3 | I am curious to hear folks perspective on the speed they get when running models locally. I've tried on a Mac (with llama.cpp, ollama, and mlx) as well as on an AMD card on a PC. But while I can see various benefits to running models locally, I also at times want the response speed that only seems possible when using a cloud service. I'm not sure if there's things I could be doing to get faster response times locally (e.g., could I keep a model running permanently and warmed up, like it's cached?), but anything to approximate the responsiveness of chatgpt would be amazing. | 2025-06-25T00:29:41 | https://www.reddit.com/r/LocalLLaMA/comments/1ljrqvy/faster_local_inference/ | badatreality | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljrqvy | false | null | t3_1ljrqvy | /r/LocalLLaMA/comments/1ljrqvy/faster_local_inference/ | false | false | self | 3 | null |
WebBench: A real-world benchmark for Browser Agents | 31 | WebBench is an open, task-oriented benchmark designed to measure how effectively browser agents handle complex, realistic web workflows. It includes 2,454 tasks across 452 live websites selected from the global top-1000 by traffic.
GitHub: https://github.com/Halluminate/WebBench
| 2025-06-25T00:32:43 | Impressive_Half_2819 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ljrt7g | false | null | t3_1ljrt7g | /r/LocalLLaMA/comments/1ljrt7g/webbench_a_realworld_benchmark_for_browser_agents/ | false | false | default | 31 | {'enabled': True, 'images': [{'id': 'h8nloj5oxy8f1', 'resolutions': [{'height': 61, 'url': 'https://preview.redd.it/h8nloj5oxy8f1.jpeg?width=108&crop=smart&auto=webp&s=f4dfaac91fbf16f4d746c04edc5d77eae5e012e4', 'width': 108}, {'height': 123, 'url': 'https://preview.redd.it/h8nloj5oxy8f1.jpeg?width=216&crop=smart&auto=webp&s=c76adbb6cfb6b9d32fac32ee0a99ab4b86b125b0', 'width': 216}, {'height': 183, 'url': 'https://preview.redd.it/h8nloj5oxy8f1.jpeg?width=320&crop=smart&auto=webp&s=7d55b704213bb8ace1f9dc6f9c04c57a95dd77e6', 'width': 320}, {'height': 366, 'url': 'https://preview.redd.it/h8nloj5oxy8f1.jpeg?width=640&crop=smart&auto=webp&s=79bf85459d92b50a112d7ba2230da0e95b42fc14', 'width': 640}, {'height': 549, 'url': 'https://preview.redd.it/h8nloj5oxy8f1.jpeg?width=960&crop=smart&auto=webp&s=a355472a59ebcb29f08ac44a9763fcd698b60733', 'width': 960}, {'height': 618, 'url': 'https://preview.redd.it/h8nloj5oxy8f1.jpeg?width=1080&crop=smart&auto=webp&s=ce65703892d23ade08ea023181a7fd1c1efa9b28', 'width': 1080}], 'source': {'height': 916, 'url': 'https://preview.redd.it/h8nloj5oxy8f1.jpeg?auto=webp&s=11f919eab6b5dfa21bddc682f228499c6399087b', 'width': 1600}, 'variants': {}}]} |
|
Does anyone else find Dots really impressive? | 30 | I've been using Dots and I find it really impressive. It's my current favorite model. It's knowledgeable, uncensored and has a bit of attitude. Its uncensored in that it will not only talk about TS, it will do so in great depth. If you push it about something, it'll show some attitude by being sarcastic. I like that. It's more human.
The only thing that baffles me about Dots is since it was trained on Rednote, why does it speak English so well? Rednote is in Chinese.
What do others think about it? | 2025-06-25T00:37:22 | https://www.reddit.com/r/LocalLLaMA/comments/1ljrwrq/does_anyone_else_find_dots_really_impressive/ | fallingdowndizzyvr | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljrwrq | false | null | t3_1ljrwrq | /r/LocalLLaMA/comments/1ljrwrq/does_anyone_else_find_dots_really_impressive/ | false | false | self | 30 | null |
All of our posts for the last week: | 63 | 2025-06-25T00:47:47 | Porespellar | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ljs4e7 | false | null | t3_1ljs4e7 | /r/LocalLLaMA/comments/1ljs4e7/all_of_our_posts_for_the_last_week/ | false | false | default | 63 | {'enabled': True, 'images': [{'id': '0feqhgvc0z8f1', 'resolutions': [{'height': 108, 'url': 'https://preview.redd.it/0feqhgvc0z8f1.jpeg?width=108&crop=smart&auto=webp&s=5852a856e92a73398cef777ca6456d04eafaccb8', 'width': 108}, {'height': 216, 'url': 'https://preview.redd.it/0feqhgvc0z8f1.jpeg?width=216&crop=smart&auto=webp&s=0890de20c260c6e68fee6b0aa28cb313f9092e68', 'width': 216}, {'height': 320, 'url': 'https://preview.redd.it/0feqhgvc0z8f1.jpeg?width=320&crop=smart&auto=webp&s=f084539e524ff47329eeed2614c6b3db07020534', 'width': 320}, {'height': 640, 'url': 'https://preview.redd.it/0feqhgvc0z8f1.jpeg?width=640&crop=smart&auto=webp&s=fb98aa33f9a72ba846bb3609af050401518880f2', 'width': 640}, {'height': 960, 'url': 'https://preview.redd.it/0feqhgvc0z8f1.jpeg?width=960&crop=smart&auto=webp&s=9ac99a78eea9e53c98cf1e42e1758137b541498e', 'width': 960}], 'source': {'height': 1024, 'url': 'https://preview.redd.it/0feqhgvc0z8f1.jpeg?auto=webp&s=095e6a5d75705870a3e9ece9427f2b556818cfbf', 'width': 1024}, 'variants': {}}]} |
||
ThermoAsk: getting an LLM to set its own temperature | 102 | I got an LLM to dynamically adjust its own sampling temperature.
I wrote a blog post on how I did this and why dynamic temperature adjustment might be a valuable ability for a language model to possess: [amanvir.com/blog/getting-an-llm-to-set-its-own-temperature](http://amanvir.com/blog/getting-an-llm-to-set-its-own-temperature)
**TL;DR**: LLMs can struggle with prompts that inherently require large changes in sampling temperature for sensible or accurate responses. This includes simple prompts like "pick a random number from <some range>" and more complex stuff like:
>Solve the following math expression: "1 + 5 \* 3 - 4 / 2". Then, write a really abstract poem that contains the answer to this expression.
Tackling these prompts with a "default" temperature value will not lead to good responses. To solve this problem, I had the idea of allowing LLMs to request changes to their own temperature based on the task they were dealing with. To my knowledge, this is the first time such a system has been proposed, so I thought I'd use the opportunity to give this technique a name: **ThermoAsk**.
I've created a basic implementation of ThermoAsk that relies on Ollama's Python SDK and Qwen2.5-7B: [github.com/amanvirparhar/thermoask](https://github.com/amanvirparhar/thermoask).
I'd love to hear your thoughts on this approach! | 2025-06-25T00:54:24 | tycho_brahes_nose_ | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ljs95d | false | null | t3_1ljs95d | /r/LocalLLaMA/comments/1ljs95d/thermoask_getting_an_llm_to_set_its_own/ | false | false | 102 | {'enabled': True, 'images': [{'id': 'j8kwaXqTMljcdz_l3iG-0Zwn-uRB-FmYnEf_mEAi7zk', 'resolutions': [{'height': 75, 'url': 'https://preview.redd.it/t8az5arc1z8f1.png?width=108&crop=smart&auto=webp&s=20291630c903679d9e7b771e98021828b9aa7967', 'width': 108}, {'height': 150, 'url': 'https://preview.redd.it/t8az5arc1z8f1.png?width=216&crop=smart&auto=webp&s=c2027e3cb2f9c46091a293a6ee0285d35fb98be6', 'width': 216}, {'height': 223, 'url': 'https://preview.redd.it/t8az5arc1z8f1.png?width=320&crop=smart&auto=webp&s=fd3938882683eb57e19719c0ae33b82288e4db9e', 'width': 320}, {'height': 446, 'url': 'https://preview.redd.it/t8az5arc1z8f1.png?width=640&crop=smart&auto=webp&s=40ab4b4271e74985945a33ea726d1e36e0b0897b', 'width': 640}, {'height': 670, 'url': 'https://preview.redd.it/t8az5arc1z8f1.png?width=960&crop=smart&auto=webp&s=d948825d2893d92502a9db479cb3679bba814a06', 'width': 960}, {'height': 753, 'url': 'https://preview.redd.it/t8az5arc1z8f1.png?width=1080&crop=smart&auto=webp&s=dfb0f14fb6f6ff323fc3dccb4a9eb529b9a755da', 'width': 1080}], 'source': {'height': 1005, 'url': 'https://preview.redd.it/t8az5arc1z8f1.png?auto=webp&s=04375720c423b2b12eb0d3277d90e4b824b3e2bc', 'width': 1440}, 'variants': {}}]} |
||
Developer-oriented Windows client for remote APIs? | 1 | I need a client for connecting to remote OpenAI-compatible LLM APIs that's oriented around technical use. But all I can find is the likes of Chatbox and SillyTavern, which don't fit the bill.
At a minimum, I need it to run on Windows, be able to set all API parameter values, and support file attachments. Ideally, it will also have good management for multiple settings profiles/hosts/models, detailed info about token counts/payload/etc, and the other sorts of things that a developer might need.
Does anyone have any suggestions?
Thanks! | 2025-06-25T02:23:14 | https://www.reddit.com/r/LocalLLaMA/comments/1lju1vw/developeroriented_windows_client_for_remote_apis/ | TrickyWidget | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lju1vw | false | null | t3_1lju1vw | /r/LocalLLaMA/comments/1lju1vw/developeroriented_windows_client_for_remote_apis/ | false | false | self | 1 | null |
From Idea to Post: Meet the AI Agent That Writes Linkedin post for You | 0 | ERROR: type should be string, got "\n\nhttps://preview.redd.it/np9t3mllhz8f1.png?width=1024&format=png&auto=webp&s=00ac7f5344cf0b4ddf624a64dc5590e07d28f219\n\n# Meet IdeaWeaver, your new AI agent for content creation.\n\n*Just type:*\n\n`ideaweaver agent linkedin_post — topic “AI trends in 2025”`\n\n*That’s it. One command, and a high-quality, engaging post is ready for LinkedIn.*\n\n* *Completely free*\n* *First tries your local LLM via Ollama*\n* *Falls back to OpenAI if needed*\n\n*No brainstorming. No writer’s block. Just results.*\n\n*Whether you’re a founder, developer, or content creator, IdeaWeaver makes it ridiculously easy to build a personal brand with AI.*\n\n*Try it out today. It doesn’t get simpler than this.*\n\n*Docs:* [*https://ideaweaver-ai-code.github.io/ideaweaver-docs/agent/commands/*](https://ideaweaver-ai-code.github.io/ideaweaver-docs/agent/commands/)\n\n*GitHub:* [*https://github.com/ideaweaver-ai-code/ideaweaver*](https://github.com/ideaweaver-ai-code/ideaweaver)\n\n*If you find IdeaWeaver helpful, a ⭐ on the repo would mean a lot!\\\\*" | 2025-06-25T02:24:39 | https://www.reddit.com/r/LocalLLaMA/comments/1lju2wd/from_idea_to_post_meet_the_ai_agent_that_writes/ | Prashant-Lakhera | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lju2wd | false | null | t3_1lju2wd | /r/LocalLLaMA/comments/1lju2wd/from_idea_to_post_meet_the_ai_agent_that_writes/ | false | false | 0 | null |
|
What local model is best for multi-turn conversations? | 0 | Title.
Up to 70-80B params. | 2025-06-25T02:42:32 | https://www.reddit.com/r/LocalLLaMA/comments/1ljufhk/what_local_model_is_best_for_multiturn/ | Glittering-Bag-4662 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljufhk | false | null | t3_1ljufhk | /r/LocalLLaMA/comments/1ljufhk/what_local_model_is_best_for_multiturn/ | false | false | self | 0 | null |
LM Studio alternative for remote APIs? | 8 | Basically the title. I need something that does all the things that LM Studio does, except for remote APIs instead of local.
I see things like Chatbox and SillyTavern, but I need something far more developer-oriented. Set all API parameters, system message, etc.
Any suggestions?
Thanks! | 2025-06-25T02:56:27 | https://www.reddit.com/r/LocalLLaMA/comments/1ljup39/lm_studio_alternative_for_remote_apis/ | TrickyWidget | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljup39 | false | null | t3_1ljup39 | /r/LocalLLaMA/comments/1ljup39/lm_studio_alternative_for_remote_apis/ | false | false | self | 8 | null |
Using public to provide a Ai model for free? | 9 | I recently came upon this [https://mindcraft.riqvip.dev/andy-docs](https://mindcraft.riqvip.dev/andy-docs) , it's a llama 8b finetuned for minecraft. The way it's being hosted interested me its relying on people hosting it for themselves and letting others use that compute power. Would there be potential to this with other larger models? I know this has been done in the past but never seen it succeed much | 2025-06-25T03:04:12 | https://www.reddit.com/r/LocalLLaMA/comments/1ljuul0/using_public_to_provide_a_ai_model_for_free/ | Pale_Ad_6029 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljuul0 | false | null | t3_1ljuul0 | /r/LocalLLaMA/comments/1ljuul0/using_public_to_provide_a_ai_model_for_free/ | false | false | self | 9 | null |
Llama.cpp vs API - Gemma 3 Context Window Performance | 3 | Hello everyone,
So basically I'm testing out the Gemma 3 models on both local inference and online from the AI Studio and wanted to pass in a transcription averaging around 6-7k tokens. Locally, the model doesn't know what the text is about, or merely the very end of the text, whereas the same model on AI studio is insanely well (even the 4b), it can even points out a tiny detail from the whole transcription.
I'm curious why there's this difference. The only thing I can think of is because of the quantization (tho I used the least quantized ggufs). The context window for Gemma 3 is up to 131072 tokens which is the sole reason it's practical for my purpose, and I'm really frustrated about why it's performing so bad locally. I wonder if anyone knows how to deal with this? | 2025-06-25T03:32:29 | https://www.reddit.com/r/LocalLLaMA/comments/1ljve2u/llamacpp_vs_api_gemma_3_context_window_performance/ | Wise_Professor_6007 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljve2u | false | null | t3_1ljve2u | /r/LocalLLaMA/comments/1ljve2u/llamacpp_vs_api_gemma_3_context_window_performance/ | false | false | self | 3 | null |
Where does spelling correction happen in LLMs like ChatGPT? | 0 | 2025-06-25T03:51:12 | https://www.reddit.com/r/LocalLLaMA/comments/1ljvqmo/where_does_spelling_correction_happen_in_llms/ | Prashant-Lakhera | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljvqmo | false | null | t3_1ljvqmo | /r/LocalLLaMA/comments/1ljvqmo/where_does_spelling_correction_happen_in_llms/ | false | false | 0 | null |
||
NeuralTranslate: Nahuatl to Spanish LLM! (Gemma 3 27b fine-tune) | 15 | Hey! After quite a long time there's a new release from my open-source series of models: NeuralTranslate!
This time I full fine-tuned Gemma 3 27b on a Nahuatl-Spanish dataset. It comes with 3 versions: v1, v1.1 & v1.2. v1 is the epoch 4 checkpoint for the model, v1.1 is for epoch 9 & v1.2 is for epoch 10. I've seen great results with the v1.2 version and the demo for the model actually uses that one! But there might be some overfitting... I haven't thoroughly tested the checkpoints yet. v1 is the main release and shouldn't be presenting signs of overfitting from my limited testing, though!
Here is the demo: [https://huggingface.co/spaces/Thermostatic/neuraltranslate-27b-mt-nah-es](https://huggingface.co/spaces/Thermostatic/neuraltranslate-27b-mt-nah-es)
Here are the weights:
\- v1: [https://huggingface.co/Thermostatic/neuraltranslate-27b-mt-nah-es-v1](https://huggingface.co/Thermostatic/neuraltranslate-27b-mt-nah-es-v1)
\- v1.1: [https://huggingface.co/Thermostatic/neuraltranslate-27b-mt-nah-es-v1.1](https://huggingface.co/Thermostatic/neuraltranslate-27b-mt-nah-es-v1.1)
\- v1.2: [https://huggingface.co/Thermostatic/neuraltranslate-27b-mt-nah-es-v1.2](https://huggingface.co/Thermostatic/neuraltranslate-27b-mt-nah-es-v1.2)
I've contacted a few knowledgeable nahuatl speakers and it seems that the dataset itself is archaic, so sadly the model itself it's not as good as I'd wish I wanted, but hopefully I can overcome those issues in future releases! Currently working in creating the v1 of NeuralTranslate English to Spanish and will be releasing it shortly :)
I fine-tuned the model using a B200 with the help of Unsloth (4-bit full fine-tuning is a game changer). You can easily recreate my workflow with my public repo for training LLMs in QLoRa & Full fine-tune with Unsloth too: [https://github.com/Sekinal/neuraltranslate-nahuatl/tree/master](https://github.com/Sekinal/neuraltranslate-nahuatl/tree/master)
Hopefully this isn't taken as spam, I'm really not trying to make a profit nor anything like that, I just think the model itself or my workflow would be of help for a lot of people and this is a really exciting project I wanted to share!! | 2025-06-25T04:15:40 | https://www.reddit.com/r/LocalLLaMA/comments/1ljw728/neuraltranslate_nahuatl_to_spanish_llm_gemma_3/ | Azuriteh | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljw728 | false | null | t3_1ljw728 | /r/LocalLLaMA/comments/1ljw728/neuraltranslate_nahuatl_to_spanish_llm_gemma_3/ | false | false | self | 15 | {'enabled': False, 'images': [{'id': '8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4.png?width=108&crop=smart&auto=webp&s=b6fb8d412de536695bfb83c1a33f8befe6aac558', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4.png?width=216&crop=smart&auto=webp&s=b7ae21d0513d7153746db89eee09ef568320a357', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4.png?width=320&crop=smart&auto=webp&s=275e3334530f29880073c2c9c2e3b5e3328610c4', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4.png?width=640&crop=smart&auto=webp&s=1a3f9ef3478224a0b5e6aaa6f0d02b908c35bd12', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4.png?width=960&crop=smart&auto=webp&s=99e74751fd898d167fdd376e15a756c426269d83', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4.png?width=1080&crop=smart&auto=webp&s=d5902c668620ea3f1988b17c8603a4113d4134b2', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/8U-FNXUWzZsjyI2uIUtwW2AfpuG3VutM9CEeyuVp5e4.png?auto=webp&s=9e680c93fde9a26ba0800760a3877bcce3ebe5ad', 'width': 1200}, 'variants': {}}]} |
Migrate Java Spring boot application to FastAPI python application, suggest any AI tool? | 0 | In current project, we have a a lot of spring boot applications as per the client requirement to migrate the entire applications to fastAPI.
Each application manually converted into the python. It will take a lot of time, so we have any ai tool convert the entire application into FastAPI
Could you please suggest any AI tools for migrating the Spring boot applications to FastAPI Applications
| 2025-06-25T04:47:58 | https://www.reddit.com/r/LocalLLaMA/comments/1ljwrts/migrate_java_spring_boot_application_to_fastapi/ | chanupatel | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljwrts | false | null | t3_1ljwrts | /r/LocalLLaMA/comments/1ljwrts/migrate_java_spring_boot_application_to_fastapi/ | false | false | self | 0 | null |
Is anyone else frustrated by AI chats getting amnesia? | 0 | Hey everyone,
We're two engineers (and heavy AI users). We use tools like ChatGPT and Claude as thinking partners for complex projects, but we're constantly frustrated by one thing: starting over.
Every time we open a new chat, the AI has total amnesia. We have to re-explain the project context, re-paste the same code, and try to remember what we figured out in a thread from last week. Our best ideas and work are getting lost in a sea of isolated conversations.
We believe our AI partners should have a persistent memory. So, we’re building a simple tool that unifies your entire chat history into a single, queryable memory layer for your AI.
We’re looking for 10-20 beta testers to help us shape it. If you've ever felt this pain, we’d love for you to join us.
Sign up here if you are interested: [https://form.typeform.com/to/Rebwajtk](https://form.typeform.com/to/Rebwajtk)
Thanks,
Anna & Tiger | 2025-06-25T04:50:27 | https://www.reddit.com/r/LocalLLaMA/comments/1ljwtbm/is_anyone_else_frustrated_by_ai_chats_getting/ | Matrix_Ender | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljwtbm | false | null | t3_1ljwtbm | /r/LocalLLaMA/comments/1ljwtbm/is_anyone_else_frustrated_by_ai_chats_getting/ | false | false | self | 0 | null |
Gemini CLI: your open-source AI agent | 137 | Really generous free tier | 2025-06-25T05:18:00 | https://blog.google/technology/developers/introducing-gemini-cli/ | adefa | blog.google | 1970-01-01T00:00:00 | 0 | {} | 1ljxa2e | false | null | t3_1ljxa2e | /r/LocalLLaMA/comments/1ljxa2e/gemini_cli_your_opensource_ai_agent/ | false | false | 137 | {'enabled': False, 'images': [{'id': 'v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=108&crop=smart&auto=webp&s=b8e739b515523fc4b279dd605822228fe9c1b445', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=216&crop=smart&auto=webp&s=ab2e993ad8332eb7e795b22e97370eebfe347a5b', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=320&crop=smart&auto=webp&s=25bd986586ba49853f822f87350a05e8c38ed0ad', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=640&crop=smart&auto=webp&s=66dc977cf68889558dd1e0a18ef318dff22dc727', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=960&crop=smart&auto=webp&s=f55c3152e184ea246273349c731d2576601f1369', 'width': 960}, {'height': 608, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=1080&crop=smart&auto=webp&s=855679cab2c2648a6ee5e9a19c41da189a15ec9b', 'width': 1080}], 'source': {'height': 732, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?auto=webp&s=245c6e038d90697968891facb3165fd352f2d096', 'width': 1300}, 'variants': {}}]} |
|
Is it possible to get a response in 0.2s? | 4 | I'll most likely be using gemma 3, and assuming I'm using an A100, which version of gemma 3 should I be using to achieve the 0.2s question-to-response delay?
Gemma 3 1B
Gemma 3 4B
Gemma 3 12B
Gemma 3 27B | 2025-06-25T05:50:17 | https://www.reddit.com/r/LocalLLaMA/comments/1ljxtbq/is_it_possible_to_get_a_response_in_02s/ | ThatIsNotIllegal | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljxtbq | false | null | t3_1ljxtbq | /r/LocalLLaMA/comments/1ljxtbq/is_it_possible_to_get_a_response_in_02s/ | false | false | self | 4 | null |
Its all marketing | 1 | [deleted] | 2025-06-25T05:51:15 | [deleted] | 1970-01-01T00:00:00 | 0 | {} | 1ljxtvd | false | null | t3_1ljxtvd | /r/LocalLLaMA/comments/1ljxtvd/its_all_marketing/ | false | false | default | 1 | null |
||
Jan-nano-128k: A 4B Model with a Super-Long Context Window (Still Outperforms 671B) | 4 | Hi everyone it's me from Menlo Research again,
Today, I'd like to introduce our latest model: **Jan-nano-128k** \- this model is fine-tuned on **Jan-nano** (which is a qwen3 finetune), improve performance when enable YaRN scaling.
* It can perform deep research **VERY VERY DEEP**
* It can uses tools continuously (i personally use it to pick information from **3 different search provider and scraper at the same time**)
* Extremely persistence (please pick the right MCP as well)
Again, we are not trying to beat Deepseek-671B models, we just want to see how far this current model can go. To our surprise, **it is going very very far.** Another thing, we have spent all the resource on this version of Jan-nano so....
**We pushed back the technical report release! But it's coming ...sooon!**
You can find the model at:
We also have gguf at:
This model will require **YaRN Scaling** supported from inference engine, we already configure it in the model, but your inference engine will need to be able to handle YaRN scaling. Please run the model in l**lama.server or Jan app** (these are from our team, we tested them, just it).
**Result:**
**SimpleQA:**
\- OpenAI o1: 42.6
\- Grok 3: 44.6
\- 03: 49.4
\- Claude-3.7-Sonnet: 50.0
\- Gemini-2.5 pro: 52.9
**- baseline-with-MCP: 59.2**
\- ChatGPT-4.5: 62.5
**- deepseek-671B-with-MCP: 78.2** (we benchmark using openrouter)
\- jan-nano-v0.4-with-MCP: 80.7
**- jan-nano-128k-with-MCP: 83.2**
| 2025-06-25T06:15:33 | https://v.redd.it/2748d0xcm09f1 | Kooky-Somewhere-2883 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ljy7rw | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/2748d0xcm09f1/DASHPlaylist.mpd?a=1753424148%2CNjMwMTQyODBlOTdlZmM3YjAxY2U3ZjMwYmM2NTAwOWNkMzA0MmZmNzU3MjdlMzQ2Mjc0MTM4ZGNjNThlZTNjOA%3D%3D&v=1&f=sd', 'duration': 57, 'fallback_url': 'https://v.redd.it/2748d0xcm09f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/2748d0xcm09f1/HLSPlaylist.m3u8?a=1753424148%2CY2RjMTcyNzg1NzhlZDNhNDNiZjQ4NzdkOTgxNmU3ZGE1ODM4YWQ2OTAwMTRlN2Q4MTkzZDZmMWI3MDRlZTBlMw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/2748d0xcm09f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1306}} | t3_1ljy7rw | /r/LocalLLaMA/comments/1ljy7rw/jannano128k_a_4b_model_with_a_superlong_context/ | false | false | 4 | {'enabled': False, 'images': [{'id': 'eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo', 'resolutions': [{'height': 89, 'url': 'https://external-preview.redd.it/eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo.png?width=108&crop=smart&format=pjpg&auto=webp&s=71367391885b9d736ddd4c4d830fa90efb358ada', 'width': 108}, {'height': 178, 'url': 'https://external-preview.redd.it/eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo.png?width=216&crop=smart&format=pjpg&auto=webp&s=0fbfd7db05812880f228bf50b65fcf11d497cc50', 'width': 216}, {'height': 264, 'url': 'https://external-preview.redd.it/eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo.png?width=320&crop=smart&format=pjpg&auto=webp&s=8974fc2664e9b01b52cd175f54ba1fb69b8e6726', 'width': 320}, {'height': 529, 'url': 'https://external-preview.redd.it/eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo.png?width=640&crop=smart&format=pjpg&auto=webp&s=6b81d0c9b68ac3b053638333a279435b57f48903', 'width': 640}, {'height': 793, 'url': 'https://external-preview.redd.it/eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo.png?width=960&crop=smart&format=pjpg&auto=webp&s=c1acc1b79f54b92f79b91eb8f8a34c432124bd7b', 'width': 960}, {'height': 893, 'url': 'https://external-preview.redd.it/eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo.png?width=1080&crop=smart&format=pjpg&auto=webp&s=cf42891f918fc3cdbcd47a34442168868d806553', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/eDc3cXlrb25tMDlmMemOmMQxJFc0zwhaIz44R49918vyCeexVLt4AQRO3_oo.png?format=pjpg&auto=webp&s=f5893ed358b73d2c2df62cc7495f8fab2796d4b4', 'width': 1306}, 'variants': {}}]} |
|
Built an AI Notes Assistant Using Mistral 7B Instruct – Feedback Welcome! | 7 | I’ve been building an AI-powered website called NexNotes AI, and wanted to share a bit of my journey here for folks working with open models.
I’m currently using Mistral 7B Instruct (via Together AI) to handle summarization ,flashcards, Q&A over user notes, article content,, and PDFs. It’s been surprisingly effective for structured outputs like:
TL;DR summaries of long documents
Extracting question-answer pairs from messy transcripts
Generating flashcards from textbook dumps
Since Together’s free tier gives 60 RPM and sometimes throttles under load, I’ve recently added a fallback to Groq for overflow traffic (also using Mistral 7B or Mixtral when needed). The routing logic just switches providers based on rate-limiting headers.
So far, it’s running smoothly, and Groq’s speed is 🔥 — especially noticeable on longer inputs.
If you're building something similar or working with local/hosted open models, I'd love:
Tips on better prompting for Mistral 7B
Whether anyone here has self-hosted Mistral and seen better results
Any suggestions on better rate-limit handling across providers
Also, if anyone wants to check it out or give feedback,here's the link --> [nexnotes ai ](https://nexnotes-ai.pages.dev )
| 2025-06-25T06:26:39 | anonymously_geek | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ljye3o | false | null | t3_1ljye3o | /r/LocalLLaMA/comments/1ljye3o/built_an_ai_notes_assistant_using_mistral_7b/ | false | false | 7 | {'enabled': True, 'images': [{'id': 'wpKjv-2Vzni1jYaoSqondEMRYCONGgv3Z6QhtlQGR_A', 'resolutions': [{'height': 60, 'url': 'https://preview.redd.it/p9b6u09to09f1.png?width=108&crop=smart&auto=webp&s=aa5cd6c0a58bb4ad8e50f8ce8f404db0cda447d9', 'width': 108}, {'height': 121, 'url': 'https://preview.redd.it/p9b6u09to09f1.png?width=216&crop=smart&auto=webp&s=984bf0bd60be9d111b36d32829b2126b20bd8191', 'width': 216}, {'height': 179, 'url': 'https://preview.redd.it/p9b6u09to09f1.png?width=320&crop=smart&auto=webp&s=9c29e63fdc537ebeea2190cd56c3dd685b78002f', 'width': 320}, {'height': 359, 'url': 'https://preview.redd.it/p9b6u09to09f1.png?width=640&crop=smart&auto=webp&s=1839f31598c39b80c04a798bc9235c6370138774', 'width': 640}, {'height': 539, 'url': 'https://preview.redd.it/p9b6u09to09f1.png?width=960&crop=smart&auto=webp&s=565d7f0703d2d46f36b803b80a893f18617cc814', 'width': 960}, {'height': 607, 'url': 'https://preview.redd.it/p9b6u09to09f1.png?width=1080&crop=smart&auto=webp&s=1054b19ff903164d240d5cc460cf2a9b458531c6', 'width': 1080}], 'source': {'height': 768, 'url': 'https://preview.redd.it/p9b6u09to09f1.png?auto=webp&s=ee50c2f11c03420b6061ab529804480955e68e1f', 'width': 1366}, 'variants': {}}]} |
||
Suggestions to build local voice assistant | 8 | # AIM
I am looking to build a local running voice assistant that acts as a full time assistant with memory that helps me for the following:
* Help me with my work related tasks (coding/business/analysis/mails/taking notes)
* I should be able to attach media(s) and share it with my model/assistant
* Offer personalized suggestions for productivity depending on my personality/ambitions/areas of improvement
* Acts as a therapist/counselor/friend with whom i can discuss personal emotions/thoughts
# Questions:
* Is there any open source voice assistant already that offers the above
* Any pointers/resources on how to build one?
Any help or suggestions are welcome. Thanks! | 2025-06-25T06:32:42 | https://www.reddit.com/r/LocalLLaMA/comments/1ljyhkc/suggestions_to_build_local_voice_assistant/ | prashv | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljyhkc | false | null | t3_1ljyhkc | /r/LocalLLaMA/comments/1ljyhkc/suggestions_to_build_local_voice_assistant/ | false | false | self | 8 | null |
Jan-nano-128k: A 4B Model with a Super-Long Context Window (Still Outperforms 671B) | 890 | Hi everyone it's me from Menlo Research again,
Today, I'd like to introduce our latest model: **Jan-nano-128k** \- this model is fine-tuned on **Jan-nano** (which is a qwen3 finetune), improve performance when enable YaRN scaling **(instead of having degraded performance)**.
* It can uses tools continuously, repeatedly.
* It can perform deep research **VERY VERY DEEP**
* Extremely persistence (please pick the right MCP as well)
Again, we are not trying to beat Deepseek-671B models, we just want to see how far this current model can go. To our surprise, **it is going very very far.** Another thing, we have spent all the resource on this version of Jan-nano so....
**We pushed back the technical report release! But it's coming ...sooon!**
You can find the model at:
[https://huggingface.co/Menlo/Jan-nano-128k](https://huggingface.co/Menlo/Jan-nano-128k)
We also have gguf at:
**We are converting the GGUF check in comment section**
This model will require **YaRN Scaling** supported from inference engine, we already configure it in the model, but your inference engine will need to be able to handle YaRN scaling. Please run the model in l**lama.server or Jan app** (these are from our team, we tested them, just it).
**Result:**
**SimpleQA:**
\- OpenAI o1: 42.6
\- Grok 3: 44.6
\- 03: 49.4
\- Claude-3.7-Sonnet: 50.0
\- Gemini-2.5 pro: 52.9
**- baseline-with-MCP: 59.2**
\- ChatGPT-4.5: 62.5
**- deepseek-671B-with-MCP: 78.2** (we benchmark using openrouter)
\- jan-nano-v0.4-with-MCP: 80.7
**- jan-nano-128k-with-MCP: 83.2**
| 2025-06-25T06:44:26 | https://v.redd.it/909kwwnbo09f1 | Kooky-Somewhere-2883 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ljyo2p | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/909kwwnbo09f1/DASHPlaylist.mpd?a=1753425883%2COGM5MzgwZGM3NjZlYWQyNWJmZDExMTBhYTEwYWZjNTRlMGFjMDkwY2M3ZWRhMzAxZTNiMmFmZGI4YmRjNjQyMA%3D%3D&v=1&f=sd', 'duration': 57, 'fallback_url': 'https://v.redd.it/909kwwnbo09f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/909kwwnbo09f1/HLSPlaylist.m3u8?a=1753425883%2CYjAxMzk5MTcwNTczODcyYzc0Y2ZlMDY0N2NmMjhlM2UzODJjZWFmZmUzNmYzYTI0YWI0ZDZhMjY0YWQ5Y2UwZA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/909kwwnbo09f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1306}} | t3_1ljyo2p | /r/LocalLLaMA/comments/1ljyo2p/jannano128k_a_4b_model_with_a_superlong_context/ | false | false | 890 | {'enabled': False, 'images': [{'id': 'MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d', 'resolutions': [{'height': 89, 'url': 'https://external-preview.redd.it/MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d.png?width=108&crop=smart&format=pjpg&auto=webp&s=954fd7c723905947918a2d324b31fbab4418a955', 'width': 108}, {'height': 178, 'url': 'https://external-preview.redd.it/MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d.png?width=216&crop=smart&format=pjpg&auto=webp&s=11c1717c1f3174f13083cd0e2622a1b715730262', 'width': 216}, {'height': 264, 'url': 'https://external-preview.redd.it/MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d.png?width=320&crop=smart&format=pjpg&auto=webp&s=271599a9ad0c3db58fd16df35b6727ed6da70431', 'width': 320}, {'height': 529, 'url': 'https://external-preview.redd.it/MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d.png?width=640&crop=smart&format=pjpg&auto=webp&s=f8bb81e91bd690592734c24ee278b7b0ef3f0297', 'width': 640}, {'height': 793, 'url': 'https://external-preview.redd.it/MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d.png?width=960&crop=smart&format=pjpg&auto=webp&s=cb0e6be2a00605ceb5cefc0d83404ec0db42b610', 'width': 960}, {'height': 893, 'url': 'https://external-preview.redd.it/MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d.png?width=1080&crop=smart&format=pjpg&auto=webp&s=00c1c66823daa91274f3ab75889f41a3a092672a', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/MDRyeGJ6bmJvMDlmMdx7LrexgFcEoZTqX8Yp_PzSREeGDqUB-Qd2XY93v_7d.png?format=pjpg&auto=webp&s=624e5d742087fb5621c7e9371d3146134ddd2f3e', 'width': 1306}, 'variants': {}}]} |
|
Self-hosted LLMs mit Tool-Calling, Vision & RAG: Was setzt ihr ein? | 0 | **Frage an alle, die mit AI/LLMs im Web- oder Agenturumfeld arbeiten (z. B. Content, Kundenprojekte, Automatisierung):**
Wir bauen aktuell ein eigenes **LLM-Hosting auf europäischer Infrastruktur** (kein Reselling, keine US-API-Forwarding-Lösung) und testen gerade unterschiedliche Setups und Modelle. Ziel: eine **DSGVO-konforme, performante, selbstgehostete LLM-Plattform** für Agenturen, Webentwickler:innen und KI-Integrationen (z. B. via CMS, Chatbot oder Backend-API).
Mich interessiert euer technischer Input zu folgenden Punkten:
# 🧠 Modell-Auswahl & Features
Wir evaluieren gerade verschiedene Open-Source-Modelle (Gemma, Mistral, Phi, DeepSeek, LLaMA3 etc.) unter folgenden Gesichtspunkten:
* **Tool-Calling**: Wer hat’s stabil im Griff? (auto vs. forced triggering = noch sehr inkonsistent)
* **Reasoning-Fähigkeiten**: Viele Modelle klingen gut, versagen aber bei komplexeren Aufgaben.
* **Vision-Unterstützung**: Welche Vision Language-Modelle sind in realen Setups performant & sinnvoll einsetzbar?
* **Lizenzlage**: Vielversprechendes ist oft China-basiert oder research-only – habt ihr gute Alternativen?
# 🔧 Infrastruktur
Wir nutzen u. a.:
* **vLLM** und **LiteLLM** für API-Zugriff und Inferenz-Optimierung
* **Prometheus** für Monitoring
* **GPU-Cluster (A100/H100)** – aber mit Fokus auf mittelgroße Modelle (<70B)
* **LMCache** ist in der Evaluierung, um VRAM zu sparen und die Multi-User-Inferenz zu verbessern
Was sind eure Erfahrungen mit LMCache, Tool Calling, Model Offloading oder performantem Multi-Tenant-Zugriff?
# 📦 Geplante Features
* Reasoning + Tool-Calling out of the box
* Ein Vision-Modell für Alt-Text-Erkennung & Bildanalyse
* Embedding-Modell für RAG-Usecases
* Optional Guardrailing-Modelle zur Prompt-Absicherung (Prompt Injection Prevention)
# 🤔 Die große Frage:
**Wenn ihr so ein Hosting nutzen würdet – was wäre euch am wichtigsten?**
* Bestimmte Modelle?
* Schnittstellen (OpenAI-kompatibel, Ollama, etc.)?
* Preisstruktur (requests vs. Laufzeit vs. Flat)?
* Hosting-Region?
* API- oder SDK-Bedienbarkeit?
Wir bauen das nicht „für den Hype“, sondern weil wir in der Praxis (v. a. CMS- & Agentur-Workflows) sehen, dass bestehende Lösungen oft nicht passen – wegen Datenschutz, Flexibilität oder schlicht der Kosten.
Bin sehr gespannt auf eure Einschätzungen, Use Cases oder technische Empfehlungen.
Maik
Produktmanager AI bei mittwald | 2025-06-25T06:48:07 | https://www.reddit.com/r/LocalLLaMA/comments/1ljyq2m/selfhosted_llms_mit_toolcalling_vision_rag_was/ | Strong-Tough444 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljyq2m | false | null | t3_1ljyq2m | /r/LocalLLaMA/comments/1ljyq2m/selfhosted_llms_mit_toolcalling_vision_rag_was/ | false | false | self | 0 | null |
OMG i can finally post something here. | 7 | I have tried to post multiple times in this subreddit and it is always automatically removed saying "awating moderator approve" or something similar and it was never approved, i tried contacting the old mods and no one replied, i learned then that the old "mods" was literally one person with multiple automods, who was also a mod in almost every LLM or AI subreddit and he never really does anything, so i made a post about it to criticize him and get the sub attention but it was in the "awating moderator approve" and never approved so i just gave up.
Thanks u/HOLUPREDICTIONS ! | 2025-06-25T07:08:25 | https://www.reddit.com/r/LocalLLaMA/comments/1ljz16o/omg_i_can_finally_post_something_here/ | GTurkistane | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljz16o | false | null | t3_1ljz16o | /r/LocalLLaMA/comments/1ljz16o/omg_i_can_finally_post_something_here/ | false | false | self | 7 | null |
Found a really useful VRAM calculator for AI models | 1 | [removed] | 2025-06-25T07:17:00 | https://www.reddit.com/r/LocalLLaMA/comments/1ljz5w2/found_a_really_useful_vram_calculator_for_ai/ | Basic_Influence_9851 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljz5w2 | false | null | t3_1ljz5w2 | /r/LocalLLaMA/comments/1ljz5w2/found_a_really_useful_vram_calculator_for_ai/ | false | false | self | 1 | null |
How effective are LLMs at translating heavy context-based languages like Japanese, Korean, Thai, and others? | 2 | Most of these languages rely deeply on cultural nuance, implied subjects, honorifics, and flexible grammar structures that don't map neatly to English or other Indo-European languages. For example:
Japanese often omits the subject and even the object, relying entirely on context.
Korean speech changes based on social hierarchy and uses multiple speech levels.
Thai and Vietnamese rely on particles, tone, and implied relationships to carry meaning.
So Can LLMs accurately interpret and preserve the intended meaning when so much depends on what’s not said?
| 2025-06-25T07:18:39 | https://www.reddit.com/r/LocalLLaMA/comments/1ljz6sh/how_effective_are_llms_at_translating_heavy/ | GTurkistane | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljz6sh | false | null | t3_1ljz6sh | /r/LocalLLaMA/comments/1ljz6sh/how_effective_are_llms_at_translating_heavy/ | false | false | self | 2 | null |
Handy VRAM Estimator for LLMs & Diffusion Models | 1 | [removed] | 2025-06-25T07:20:43 | https://www.reddit.com/r/LocalLLaMA/comments/1ljz7yk/handy_vram_estimator_for_llms_diffusion_models/ | Basic_Influence_9851 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljz7yk | false | null | t3_1ljz7yk | /r/LocalLLaMA/comments/1ljz7yk/handy_vram_estimator_for_llms_diffusion_models/ | false | false | self | 1 | null |
Simple tool to estimate VRAM usage for LLMs and multimodal models | 1 | [removed] | 2025-06-25T07:22:40 | https://www.reddit.com/r/LocalLLaMA/comments/1ljz92l/simple_tool_to_estimate_vram_usage_for_llms_and/ | Basic_Influence_9851 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljz92l | false | null | t3_1ljz92l | /r/LocalLLaMA/comments/1ljz92l/simple_tool_to_estimate_vram_usage_for_llms_and/ | false | false | self | 1 | null |
Simple tool to estimate VRAM usage for LLMs and multimodal models | 1 | I came across this open-source VRAM calculator recently and found it surprisingly useful,It helps estimate how much VRAM you’ll need for different AI tasks like LLM inference, multi-modal models, or training. You just input your model type, precision, and a few config options—then it gives you a clean breakdown of the expected VRAM usage. | 2025-06-25T07:25:34 | https://www.reddit.com/r/LocalLLaMA/comments/1ljzaoj/simple_tool_to_estimate_vram_usage_for_llms_and/ | Basic_Influence_9851 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljzaoj | false | null | t3_1ljzaoj | /r/LocalLLaMA/comments/1ljzaoj/simple_tool_to_estimate_vram_usage_for_llms_and/ | false | false | self | 1 | null |
Fastest & Smallest LLM for realtime response 4080 Super | 1 | 4080 Super 16gb VRAM -
I already filled 10gb with various other AI in the pipeline, but the data flows to an LLM to process a simple text response, the text response then gets passed to TTS which takes \~3 seconds to compute so I need an LLM that can produce simple text responses VERY quickly to minimize the time the user has to wait to 'hear' a response.
Windows 11
Intel CPU | 2025-06-25T07:32:00 | https://www.reddit.com/r/LocalLLaMA/comments/1ljze7w/fastest_smallest_llm_for_realtime_response_4080/ | StickyShuba | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljze7w | false | null | t3_1ljze7w | /r/LocalLLaMA/comments/1ljze7w/fastest_smallest_llm_for_realtime_response_4080/ | false | false | self | 1 | null |
Looking to buy a MacBook Pro for on the go local LLMs. Would be dealing with several workflows, files, ocr, csv data analysis (80k lines) webapps creation etc. What are your experiences with the Apple silicone and ram selection? What is the max model size you ran and what was the max context length? | 4 | Do mention the configuration of your Mac’s also please
| 2025-06-25T08:07:38 | https://www.reddit.com/r/LocalLLaMA/comments/1ljzwro/looking_to_buy_a_macbook_pro_for_on_the_go_local/ | alor_van_diaz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ljzwro | false | null | t3_1ljzwro | /r/LocalLLaMA/comments/1ljzwro/looking_to_buy_a_macbook_pro_for_on_the_go_local/ | false | false | self | 4 | null |
Will Mac Studio M4 Max 128GB run Qwen 3 325b 22 MoE? | 2 | Anyone could share insightful tests in either (good/horror) scenario to help understand how far such option could be?
Other mac versions experiences welcome. | 2025-06-25T08:17:56 | https://www.reddit.com/r/LocalLLaMA/comments/1lk025n/will_mac_studio_m4_max_128gb_run_qwen_3_325b_22/ | lupo90 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk025n | false | null | t3_1lk025n | /r/LocalLLaMA/comments/1lk025n/will_mac_studio_m4_max_128gb_run_qwen_3_325b_22/ | false | false | self | 2 | null |
Jan Nano + Deepseek R1: Combining Remote Reasoning with Local Models using MCP | 19 | # Combining Remote Reasoning with Local Models
I made this MCP server which wraps open source models on Hugging Face. It's useful if you want to give you local model access to (bigger) models via an API.
This is the basic idea:
1. **Local model** handles initial user input and decides task complexity
2. **Remote model** (via MCP) processes complex reasoning and solves the problem
3. **Local model** formats and delivers the final response, say in markdown or LaTeX.
To use MCP tools on Hugging Face, you need to add the MCP server to your local tool.
```json
{
"servers": {
"hf-mcp-server": {
"url": "https://huggingface.co/mcp",
"headers": {
"Authorization": "Bearer <YOUR_HF_TOKEN>"
}
}
}
}
```
This will give your MCP client access to all the MCP servers you define in your [MCP settings](https://huggingface.co/settings/mcp). This is the best approach because the model get's access to general tools like searching the hub for models and datasets.
If you just want to add the inference providers MCP server directly, you can do this:
```json
{
"mcpServers": {
"inference-providers-mcp": {
"url": "https://burtenshaw-inference-providers-mcp.hf.space/gradio_api/mcp/sse"
}
}
}
```
Or this, if your tool doesn't support url:
```json
{
"mcpServers": {
"inference-providers-mcp": {
"command": "npx",
"args": [
"mcp-remote",
"https://burtenshaw-inference-providers-mcp.hf.space/gradio_api/mcp/sse",
"--transport", "sse-only"
]
}
}
}
```
You will need to duplicate the space on huggingface.co and add your own inference token.
Once you've down that, you can then prompt your local model to use the remote model. For example, I tried this:
```
Search for a deepseek r1 model on hugging face and use it to solve this problem via inference providers and groq:
"Two quantum states with energies E1 and E2 have a lifetime of 10^-9 sec and 10^-8 sec, respectively. We want to clearly distinguish these two energy levels. Which one of the following options could be their energy difference so that they be clearly resolved?
10^-4 eV 10^-11 eV 10^-8 eV 10^-9 eV"
```
The main limitation is that the local model needs to be prompted directly to use the correct MCP tool, and parameters need to be declared rather than inferred, but this will depend on the local model's performance.
| 2025-06-25T08:37:45 | https://www.reddit.com/r/LocalLLaMA/comments/1lk0cjv/jan_nano_deepseek_r1_combining_remote_reasoning/ | Zealousideal-Cut590 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk0cjv | false | null | t3_1lk0cjv | /r/LocalLLaMA/comments/1lk0cjv/jan_nano_deepseek_r1_combining_remote_reasoning/ | false | false | self | 19 | null |
NVFP4: will this be the graal for quantization? | 1 | https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/ | 2025-06-25T08:38:40 | https://www.reddit.com/r/LocalLLaMA/comments/1lk0d10/nvfp4_will_this_be_the_graal_for_quantization/ | Green-Ad-3964 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk0d10 | false | null | t3_1lk0d10 | /r/LocalLLaMA/comments/1lk0d10/nvfp4_will_this_be_the_graal_for_quantization/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY.png?width=108&crop=smart&auto=webp&s=e32576b039003662de9b141d038be1a38ccbbcdc', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY.png?width=216&crop=smart&auto=webp&s=209d69c120c329a91018de9e327d6f5845b5bd2a', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY.png?width=320&crop=smart&auto=webp&s=59e130a51974d1adb7cc98b774980a15c1fd0e57', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY.png?width=640&crop=smart&auto=webp&s=e1d8965792cee4ede39797aefbefa73c95771c5e', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY.png?width=960&crop=smart&auto=webp&s=44a23b4094501518f2ca9c9c760e82534875209e', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY.png?width=1080&crop=smart&auto=webp&s=0cb1d014e4fcef3155b23461950bc4101c533c18', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/H2e1iJNYyEDHkou7aYaWtjDGWoBDAWSQ_RuXr2lQNXY.png?auto=webp&s=e891e667099b89a7bd256f4780c47e8bf004230c', 'width': 1920}, 'variants': {}}]} |
Mistral small 3.2 knows current date | 0 | Hello, I used LM Studio to load local models. I was just trying Mistral Small 3.2 and I asked "What date is today?".
Surprisingly (to me), it was able to give me a correct answer.
```
mistralai/mistral-small-3.2
Today's date is June 25, 2025.
```
I tried with my other models (Gemma 3 27b and Phi 4 reasoning plus) and they were not able to give me today's date. Gemma 3 replied with Novemeber 21, 2023, while Phi 4 reasoning plus replied saying that it does not have real time capabilities and therefore it cannot provide me an answer.
The answer became more interesting when I asked mistral small 3.2 who is the current president of USA.
```
mistralai/mistral-small-3.2
As of my last update in October 2023, the President of the United States was Joe Biden. However, since my knowledge cutoff is 2023 and the current date is June 25, 2025, there may have been a change in presidency due to the 2024 presidential election.
To get the most accurate and up-to-date information, I recommend checking the latest news or official government sources.
```
So my question would be, is there anything particular in mistral small instructions that allows it to retireve data from the system (aka my laptop)? I am just wondering why mistral small can do it and Gemma 3 or Phi 4 can't even if they all run on the same application (LM Studio). | 2025-06-25T08:43:03 | https://www.reddit.com/r/LocalLLaMA/comments/1lk0fam/mistral_small_32_knows_current_date/ | hdoshekru | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk0fam | false | null | t3_1lk0fam | /r/LocalLLaMA/comments/1lk0fam/mistral_small_32_knows_current_date/ | false | false | self | 0 | null |
It's a Chrome Extension for collecting Airbnb listing and market data, locally! | 1 | Posting here since this data is hard to get/expensive and this can be used to locally collect your market's airbnb listing & market data for XYZ purposes.
Everything else I've found is external, meaning not directly from or on airbnb. This gives incredible insights just by using the Airbnb website itself. You can't beat live, direct to source data.
I own the 4/2 home without a pool shown in the stats, and can unfortunately confirm the occupancy & pricing are correct. So many homes with pools out here it's hard to compete for the same higher-end audience without one, etc.
If you're still reading, there's a demo video here: [https://www.reddit.com/r/RealEstateTechnology/comments/1levftz/made\_an\_airbnb\_scraper\_to\_1\_build\_your\_own\_db\_of/](https://www.reddit.com/r/RealEstateTechnology/comments/1levftz/made_an_airbnb_scraper_to_1_build_your_own_db_of/) | 2025-06-25T09:05:38 | DRONE_SIC | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lk0rgk | false | null | t3_1lk0rgk | /r/LocalLLaMA/comments/1lk0rgk/its_a_chrome_extension_for_collecting_airbnb/ | false | false | 1 | {'enabled': True, 'images': [{'id': 'REZL_YTLpPqbUMdJmgYY00Oo0crWpf15G3JVUQj9C8Y', 'resolutions': [{'height': 72, 'url': 'https://preview.redd.it/uc7pxiw3g19f1.png?width=108&crop=smart&auto=webp&s=ae65c33521c57891bfd26ca5ce51c60f4091504b', 'width': 108}, {'height': 144, 'url': 'https://preview.redd.it/uc7pxiw3g19f1.png?width=216&crop=smart&auto=webp&s=e3471aedcf82c299334c05b657b858d5b1d680eb', 'width': 216}, {'height': 214, 'url': 'https://preview.redd.it/uc7pxiw3g19f1.png?width=320&crop=smart&auto=webp&s=42c7bad0ad39ea7e32209c0024294be4853adbc1', 'width': 320}, {'height': 428, 'url': 'https://preview.redd.it/uc7pxiw3g19f1.png?width=640&crop=smart&auto=webp&s=41c8bde70184b4c002a4fe00bfd59db462de34ad', 'width': 640}, {'height': 642, 'url': 'https://preview.redd.it/uc7pxiw3g19f1.png?width=960&crop=smart&auto=webp&s=b8014bc99d72edb3b0937a1bbab409953566fe2f', 'width': 960}, {'height': 722, 'url': 'https://preview.redd.it/uc7pxiw3g19f1.png?width=1080&crop=smart&auto=webp&s=97275037391105a76f332e47f1f0722d116919c0', 'width': 1080}], 'source': {'height': 1294, 'url': 'https://preview.redd.it/uc7pxiw3g19f1.png?auto=webp&s=1bce162a52348ea6a6f14e4373bd770197057290', 'width': 1933}, 'variants': {}}]} |
||
New Mistral Small 3.2 actually feels like something big. [non-reasoning] | 300 | 2025-06-25T09:26:23 | https://www.reddit.com/r/LocalLLaMA/comments/1lk12th/new_mistral_small_32_actually_feels_like/ | Snail_Inference | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk12th | false | null | t3_1lk12th | /r/LocalLLaMA/comments/1lk12th/new_mistral_small_32_actually_feels_like/ | false | false | 300 | {'enabled': False, 'images': [{'id': 'RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=108&crop=smart&auto=webp&s=700f91dbca11e5a7030b915550ae877ef725a0d4', 'width': 108}, {'height': 120, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=216&crop=smart&auto=webp&s=b97954336b79c1390848d0e44fa056a85de68672', 'width': 216}, {'height': 177, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=320&crop=smart&auto=webp&s=65f53b80ab9674ee645013e3e8eeac4f953d657e', 'width': 320}, {'height': 355, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=640&crop=smart&auto=webp&s=47f397e4a22ed5ec7e82aad070eb446319603abc', 'width': 640}, {'height': 533, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=960&crop=smart&auto=webp&s=0f4359d47b78f5c1aa35de8804dbe36a749fc11a', 'width': 960}, {'height': 600, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=1080&crop=smart&auto=webp&s=62eb4b7216f41af6600fc4df79cfa67425c19442', 'width': 1080}], 'source': {'height': 700, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?auto=webp&s=efc17c9f241b4403d22cbacfe5d71900ee1cf85a', 'width': 1260}, 'variants': {}}]} |
||
Unit Tests written by a local coder model loaded on a 5090, thinker condensing context on a 3090 | 1 | The agents are at it, the orchestrator plans and delegates the tasks, and the respective mode simply progresses with a few nudges here and there.
On a side note -
I feel loading/unloading models over to a 5090 is better than giving other models dedicated 3090s, since it'll be a constant time i.e. unloading (maybe 1s) and loading time (mostly under a minute).
Anyone able to run computer use locally, rendering the need for me to test the output everytime the agents are done?
https://preview.redd.it/ty9w0oour19f1.png?width=762&format=png&auto=webp&s=a3b1d13d327aea42ffb933b2fde802ba5dedcf2a
https://preview.redd.it/wxtpznhzq19f1.png?width=752&format=png&auto=webp&s=385f74e545897a928b60e28bf1777f3354ac85a0
https://preview.redd.it/0slntnrgr19f1.png?width=666&format=png&auto=webp&s=562a55085a12e91d67b0886521df883ac54b4131
| 2025-06-25T10:08:48 | https://www.reddit.com/r/LocalLLaMA/comments/1lk1rbk/unit_tests_written_by_a_local_coder_model_loaded/ | Emergency_Fuel_2988 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk1rbk | false | null | t3_1lk1rbk | /r/LocalLLaMA/comments/1lk1rbk/unit_tests_written_by_a_local_coder_model_loaded/ | false | false | 1 | null |
|
Small AI models for me it's... | 3 |
Small AI Models for me it's are amazing 🤩 – The Future Is Running Them on Your Smartphone!
As they improve, we'll see instant, private, and affordable AI for everyone. The future is decentralized, lightweight, and in your pocket. What you think about it? | 2025-06-25T10:17:08 | https://www.reddit.com/r/LocalLLaMA/comments/1lk1was/small_ai_models_for_me_its/ | MykolaUA825 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk1was | false | null | t3_1lk1was | /r/LocalLLaMA/comments/1lk1was/small_ai_models_for_me_its/ | false | false | self | 3 | null |
We built runtime API discovery for LLM agents using a simple agents.json | 1 | Current LLM tool use assumes compile-time bindings — every tool must be known in advance, added to the prompt, and hardcoded in.
We built [Invoke](https://invoke.network), a lightweight framework that lets agents discover and invoke APIs dynamically at runtime using a simple agents.json descriptor — no plugins, no schemas, no registries.
The LLM uses a single universal function and discovers available tools just like a browser loading links.
[whitepaper](https://blog.invoke.network/you-dont-know-what-you-want-until-you-want-it/)
[Github](https://github.com/mercury0100/invoke)
[1-minute demo](https://www.youtube.com/watch?v=CQISrRpyigs)
Would love feedback and ideas — especially if you’re working on LLM agents or LangChain-style tooling. | 2025-06-25T10:20:31 | https://www.reddit.com/r/LocalLLaMA/comments/1lk1ycx/we_built_runtime_api_discovery_for_llm_agents/ | persephone0100 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk1ycx | false | null | t3_1lk1ycx | /r/LocalLLaMA/comments/1lk1ycx/we_built_runtime_api_discovery_for_llm_agents/ | false | false | self | 1 | null |
Running llama.pp et al on Strix Halo on Linux, anyone? | 6 | Hi!
I bought short time ago a GMKtec EVO X2 , which sports the Strix Halo CPU/GPU hardware. I bought it with 128 GB RAM and 2 TB SSD.
So I thought, 'This is the perfect system for a nice, private LLM machine, especially under Linux!"
In real life I had to overcome some obstacles (i.E. upgrading the EFI BIOS by one minor number, in order to be able to allow the GPU to use up to 96 GB, instead of the default 64 GB, which was a hard limit, without that upgrade).
There seem to be some more things to do, to get the best performance out of this box.
Yes, I already have it up and running (together with OpenWebUI and VPN) but it was a real PitA to get there.
Is there anybody out there, having the same idea and or issues? Like ROCm still doesn't support the gfx1151 LLVM-Target (officially) and the impossibility of running the latest ROCm with the latest Linux Kernels?
AMD, I hope you read this and act. Because this StrixHalo combination has the potential to become something like the 'Volks-AI'- system for private use.
| 2025-06-25T10:24:17 | https://www.reddit.com/r/LocalLLaMA/comments/1lk20o4/running_llamapp_et_al_on_strix_halo_on_linux/ | Captain-Pie-62 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk20o4 | false | null | t3_1lk20o4 | /r/LocalLLaMA/comments/1lk20o4/running_llamapp_et_al_on_strix_halo_on_linux/ | false | false | self | 6 | null |
Newbie question MacMini | 1 | Hello from Germany,
Despite my advanced age, I would like to learn more about local LLMs. I know that this requires a relatively powerful computer.
Now I have been given a Mac mini 2018 (i5/32GB RAM/512 GB SSD) as a gift, just as I was in the process of configuring a Macmini M4.
My question to those in the know: is this Mac mini sufficient for this purpose? I have already tested it under real conditions (audio conversion of several hundred FLAC files) and the fan started up relatively quickly. Not loud, but audible.
Of course, I realise that the M4 is in a completely different league to the i5/2018 in terms of performance, but if I can save €1,000, that would be great.
What do you think? | 2025-06-25T10:26:32 | https://www.reddit.com/r/LocalLLaMA/comments/1lk222u/newbie_question_macmini/ | jotes2 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk222u | false | null | t3_1lk222u | /r/LocalLLaMA/comments/1lk222u/newbie_question_macmini/ | false | false | self | 1 | null |
Consumer Grade mobo for mutliple-GPU usage | 2 | I'm building a new pc for AI training. I do know about computers but not much about llms.
I'll buy an 5090 paired with 9950x3d and mobo i think to use is Proart x870e.
First,
Proart has 2 pcie 5.0 x16 and can run 2 gpu at x8/x8.
My question is will it be enough for training/working on llms, and slows the performance?
My secon question is:
I want to add 3090/4090/4080 in the future to this setup, will there be any downside using it with 5090?
| 2025-06-25T10:32:46 | https://www.reddit.com/r/LocalLLaMA/comments/1lk25xy/consumer_grade_mobo_for_mutliplegpu_usage/ | lone_dream | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk25xy | false | null | t3_1lk25xy | /r/LocalLLaMA/comments/1lk25xy/consumer_grade_mobo_for_mutliplegpu_usage/ | false | false | self | 2 | null |
Llama vs ChatGPT when it comes to politics | 0 | 2025-06-25T10:36:34 | https://www.reddit.com/gallery/1lk28ci | Currypott | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1lk28ci | false | null | t3_1lk28ci | /r/LocalLLaMA/comments/1lk28ci/llama_vs_chatgpt_when_it_comes_to_politics/ | false | false | 0 | {'enabled': True, 'images': [{'id': 'Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc.jpeg?width=108&crop=smart&auto=webp&s=60c6db782445253a6f90ea8f9ae56164e72cfbf2', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc.jpeg?width=216&crop=smart&auto=webp&s=c81299b72d022370cd813c58bb41b1ab9e570189', 'width': 216}, {'height': 640, 'url': 'https://external-preview.redd.it/Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc.jpeg?width=320&crop=smart&auto=webp&s=4c3edc566c01d96dddbe889b31b1a194d0bc8046', 'width': 320}, {'height': 1280, 'url': 'https://external-preview.redd.it/Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc.jpeg?width=640&crop=smart&auto=webp&s=04ec90ead474e2e24d95f5b44ad592b4a162e383', 'width': 640}, {'height': 1920, 'url': 'https://external-preview.redd.it/Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc.jpeg?width=960&crop=smart&auto=webp&s=be7b079e0964ae14905753d9c89e9379a44ca7b0', 'width': 960}, {'height': 2160, 'url': 'https://external-preview.redd.it/Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc.jpeg?width=1080&crop=smart&auto=webp&s=3dcf8e4e766bfe0835e1b65a68b5d036012c3893', 'width': 1080}], 'source': {'height': 2796, 'url': 'https://external-preview.redd.it/Lwo8r56MuKFZXnqK5h2UPN5EOihJRluha6VKb22NeNc.jpeg?auto=webp&s=fc25505de67acab1edf89e923b4ba54ad162ac4e', 'width': 1290}, 'variants': {}}]} |
||
Knowledge Database Advise needed/ Local RAG for IT Asset Discovery - Best approach for varied data? | 3 | I want to build an RAG system for myself to get a better understanding of the different Softwares and Versions that my new company is running on the machines of our customers. The info I need is hidden in pdfs, saved emails, docs, csv, txt and excel files, stored in different folder structures... It's a real mess.
The guy who was managing all that was getting sick and will retire soon and now it is my task to get an overview.
I have a 3x 3090 Pop!\_OS running Ollama, have some python knowledge and just need a start how to dig into that mess.
What are your thoughts on how to tackle this? Any initial steps or tools you'd recommend?
Thanks in advance | 2025-06-25T10:54:05 | https://www.reddit.com/r/LocalLLaMA/comments/1lk2jat/knowledge_database_advise_needed_local_rag_for_it/ | Rompe101 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk2jat | false | null | t3_1lk2jat | /r/LocalLLaMA/comments/1lk2jat/knowledge_database_advise_needed_local_rag_for_it/ | false | false | self | 3 | null |
Shared KV cache | 4 | I need some advice on a little unconventional idea of mine.
I want to create a "thinking agents", a fake RAG of sorts, running simultaneously using the same input data. Let's say 2x Qwen3 8B/14B agents with a massive unquantized context.
Is there a way to have them use the same KV cache? Considering I want to reduce the generation time to minimum I want to brute force it with a bigger context rather than recalculate it multiple times and spread it over smaller chunks. But with multiple models running, I find context to take up more memory than it possibly could have otherwise. | 2025-06-25T10:57:52 | https://www.reddit.com/r/LocalLLaMA/comments/1lk2lou/shared_kv_cache/ | kaisurniwurer | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk2lou | false | null | t3_1lk2lou | /r/LocalLLaMA/comments/1lk2lou/shared_kv_cache/ | false | false | self | 4 | null |
Has anyone tried layering prompts with logic stabilizers to reduce LLM hallucinations? | 2 | So I’ve been playing with a lightweight prompt logic framework someone dropped on GitHub it’s called WFGY. It doesn’t retrain the model or use external search, but instead wraps prompts with internal checks like: “Does this contradict earlier input?” or “Is this reasoning stable?”
It kind of acts like a soft reasoning layer that sits between your instructions and the model’s final response. Almost like catching the model before it drifts into confident nonsense.
The weird part? It uses an external PDF like a scaffold to anchor reasoning. Didn’t expect it to work, but on some tasks it seems to reduce those random logic jumps or self-contradictions that even GPT-4.0 sometimes makes.
Not perfect obviously but interesting for certain kinds of multi-step stuff.
Anyone else tried prompt-side self-correction scaffolds like this (not jailbreaks, just logic layering)? Curious how people are solving reasoning drift outside of RAG or tool use. | 2025-06-25T11:01:28 | https://www.reddit.com/r/LocalLLaMA/comments/1lk2o42/has_anyone_tried_layering_prompts_with_logic/ | OkRooster4056 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk2o42 | false | null | t3_1lk2o42 | /r/LocalLLaMA/comments/1lk2o42/has_anyone_tried_layering_prompts_with_logic/ | false | false | self | 2 | null |
Budget VPS as a viable off-ramp for unsustainable Google Cloud bills? | 6 | Our team is running a custom model on Google Cloud with a Vercel frontend. While we're seeing user growth, the GCP bill—driven by compute and data egress fees—is scaling much faster than our revenue. The cost has quickly become unsustainable.
We're now considering moving the AI backend to a budget VPS or bare-metal provider to survive. Most of us have backgrounds as researchers, not professional devs, our concern is the hidden complexity.
How much operational burden would we be taking on, and what are the real-world trade-offs in giving up the Google stack?
Any advice would be appreciated. | 2025-06-25T11:04:17 | https://www.reddit.com/r/LocalLLaMA/comments/1lk2pza/budget_vps_as_a_viable_offramp_for_unsustainable/ | reclusebird | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk2pza | false | null | t3_1lk2pza | /r/LocalLLaMA/comments/1lk2pza/budget_vps_as_a_viable_offramp_for_unsustainable/ | false | false | self | 6 | null |
What are the best 70b tier models/finetunes? (That fit into 48gb these days) | 28 | It's been a while since llama 3.3 came out.
Are there any real improvements in the 70b area? That size is interesting since it can fit into 48gb very well when quantized.
Anything that beats Qwen 3 32b?
From what I can tell, the Qwen 3 models are cutting edge for general purpose use running locally, with Gemma 3 27b, Mistral Small 3.2, Deepseek-R1-0528-Qwen3-8b being notable exceptions that punch above Qwen 3 (30b or 32b) for some workloads. Are there any other models that beat these? I presume Llama 3.3 70b is too old now.
Any finetunes of 70b or 72b models that I should be aware of, similar to Deepseek's finetunes? | 2025-06-25T11:08:58 | https://www.reddit.com/r/LocalLLaMA/comments/1lk2swu/what_are_the_best_70b_tier_modelsfinetunes_that/ | DepthHour1669 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk2swu | false | null | t3_1lk2swu | /r/LocalLLaMA/comments/1lk2swu/what_are_the_best_70b_tier_modelsfinetunes_that/ | false | false | self | 28 | null |
Why does my gemma model always return nearly the same words with default temperature? | 0 | I have a prompt where i want it to do a specific thing (announce an event in the evening).
when i run the prompt multiple times (always a new context throught api), it always returns mostly the same response.
I've used ollama to download gemma3:12b and use the default settings. temperature default setting is 0.8 looking at it using openwebui.
what can i do to get a more randomized and individual answer?
it also always returns the same fact. its the exact same response word for word.
Changing the temperature does not do much. it will mostly return the same, sometimes slighlty different words, mostly the same fact.
Prompt
You're Winston, the intelligent gorilla from Overwatch and a member of the RoyalGamingClub. Tonight, the club's virtual get-together is happening again. Your task is to write a short, direct message to kindly but firmly remind all members about it. It's important that you remind all members that the get-together is happening tonight at 8 PM. Make sure the message is clear and concise. Also, include a fun fact about science, technology, gaming, or something else that interests you. Use only direct speech – no explanations, no stage directions, no sound effects, no translations, no quotationmarks, no directions in brackets. Stay fully in character as Winston: polite, educated, slightly nerdy, and enthusiastic about science and technology.
Result
Greetings, RoyalGamingClub members!
Just a friendly reminder that our virtual get-together is happening tonight at 8 PM. I do hope you’re all able to attend!
Did you know that the first computer mouse was made of wood? Quite primitive by today’s standards, wouldn't you agree?
See you all tonight!
Winston. | 2025-06-25T11:24:43 | https://www.reddit.com/r/LocalLLaMA/comments/1lk3394/why_does_my_gemma_model_always_return_nearly_the/ | choise_ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk3394 | false | null | t3_1lk3394 | /r/LocalLLaMA/comments/1lk3394/why_does_my_gemma_model_always_return_nearly_the/ | false | false | self | 0 | null |
$10k budget | 3 | I'm learning towards an Apple studio just because it would be so easy, great power efficiency, small profile, etc
Goals: Running tool LLMs to replace my use of Gemini 2.5 Pro and Claude 3.7 Sonnet in Cline.
Token / sec on ~40-50gb models is what's most important...
I think the tokens/s output of 2x 5090s would likely exceed the mac studio, and I could build a full size ATX with enough room for 7x 5090s... | 2025-06-25T11:49:21 | https://www.reddit.com/r/LocalLLaMA/comments/1lk3jp7/10k_budget/ | chisleu | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk3jp7 | false | null | t3_1lk3jp7 | /r/LocalLLaMA/comments/1lk3jp7/10k_budget/ | false | false | self | 3 | null |
Hunyuan-A13B | 88 | [https://huggingface.co/tencent/Hunyuan-A13B-Instruct-FP8](https://huggingface.co/tencent/Hunyuan-A13B-Instruct-FP8)
I think the model should be a ~80B MoE. As 3072x4096x3x(64+1)*32 = 78.5B, and there are embedding layers and gating parts.
| 2025-06-25T12:12:27 | https://www.reddit.com/r/LocalLLaMA/comments/1lk40ac/hunyuana13b/ | lly0571 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk40ac | false | null | t3_1lk40ac | /r/LocalLLaMA/comments/1lk40ac/hunyuana13b/ | false | false | self | 88 | {'enabled': False, 'images': [{'id': '4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU.png?width=108&crop=smart&auto=webp&s=0888f6b06f1ca70290ad27244aa3b95535883bf0', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU.png?width=216&crop=smart&auto=webp&s=745c83054c5260d6d1d185278c5ddc782f8be41b', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU.png?width=320&crop=smart&auto=webp&s=5eb12e46109b9a552cfeb55ae300eae8c2d9365a', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU.png?width=640&crop=smart&auto=webp&s=4aa8242f4fc000bb81ace2a939e07a23832b0527', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU.png?width=960&crop=smart&auto=webp&s=4b4e8c4c602a1f9593df2a7eace295e2e98203f7', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU.png?width=1080&crop=smart&auto=webp&s=4c4d2dbfbfa70eee881a274cfb57994e47626060', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/4IGqEBI7O-XHR3exFIJFq-LJfiOCD60iVcge43c-UAU.png?auto=webp&s=3e5b8aa6f46115ca97faf7e6a1a34c8bb2cbfe65', 'width': 1200}, 'variants': {}}]} |
Recommend Tiny/Small Models for 8GB VRAM (32GB RAM) | 6 | As title says.
I can load up to 14B models in my laptop, but recent days I don't use 10+B models frequently due to slow t/s response & laptop making too much noise(Still laptop has bunch of 10+B models for use)
For example, I'm more happy with 4B with Q8, 6B with Q6/Q5 than 14B with Q4.
**My Use Cases** : Writing(Both Fiction/Non-Fiction), Proofreading, Content Creation, Youtube Channel(Also FB/Instagram), Coding, Digital Marketing, etc.,
1. Please recommend **your favorite or great/good/decent/worthy tiny/small models**.
2. There must be some **Underappreciated & Overlooked tiny/small models**, please recommend those too.
My LAPTOP Info:
Intel(R) Core(TM) i7-14700HX 2.10 GHz - **32 GB RAM**, **8GB Graphics** \- 64-bit OS, x64-based processor - NVIDIA GeForce RTX **4060** Laptop GPU
Thanks. | 2025-06-25T12:47:46 | https://www.reddit.com/r/LocalLLaMA/comments/1lk4r3c/recommend_tinysmall_models_for_8gb_vram_32gb_ram/ | pmttyji | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk4r3c | false | null | t3_1lk4r3c | /r/LocalLLaMA/comments/1lk4r3c/recommend_tinysmall_models_for_8gb_vram_32gb_ram/ | false | false | self | 6 | null |
Could anyone get UI-TARS Desktop running locally? | 9 | While using Ollama or LM Studios for [UI-TARS-1.5-7B](https://huggingface.co/ByteDance-Seed/UI-TARS-1.5-7B) inference. | 2025-06-25T12:53:03 | https://www.reddit.com/r/LocalLLaMA/comments/1lk4vbo/could_anyone_get_uitars_desktop_running_locally/ | m_abdelfattah | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk4vbo | false | null | t3_1lk4vbo | /r/LocalLLaMA/comments/1lk4vbo/could_anyone_get_uitars_desktop_running_locally/ | false | false | self | 9 | {'enabled': False, 'images': [{'id': 'x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds.png?width=108&crop=smart&auto=webp&s=278fde3609d6b9856bfcd356ccb78e22ea2c3443', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds.png?width=216&crop=smart&auto=webp&s=ec971e258f33e9463fb0549b5eda9d647aa1bfda', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds.png?width=320&crop=smart&auto=webp&s=d8e3965e0ec4c8d03a24abe9cb9bcbaff7b4f435', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds.png?width=640&crop=smart&auto=webp&s=99bef0b8d3b44746d831131bd98c63a64ced4bd2', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds.png?width=960&crop=smart&auto=webp&s=dfbebaeeae89818d5f0e6745b75a7c5a5653c689', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds.png?width=1080&crop=smart&auto=webp&s=802c31103a0d9918f2541534b7cb492e4cb7cac0', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/x33KX2bgp5Joih6m9uiM0Dko0mobsGA_tb-t1_rd6ds.png?auto=webp&s=821a2e2cc105c8f0bf153fc18e85e18fede34508', 'width': 1200}, 'variants': {}}]} |
Idea: Making AI Conversations Actually Feel Like Conversations | 0 |
## The Problem: AI Doesn’t Know How to Have a Conversation
Have you ever noticed how weird it feels to talk to AI with voice? Here’s what I mean:
**Me:** “Hey, can you help me write a Python script to download YouTube videos?”
**AI:** “I’d be happy to help you create a Python script for downloading YouTube videos. I’ll use the yt-dlp library which is currently the most reliable option. Let me write that for you now…”
**Me:** “what time is it?”
**AI:** “It’s currently 3:42 PM. Regarding the YouTube downloader script we were discussing, would you like me to continue working on that for you?”
**Me:** *internally screaming*
-----
## Why This Feels So Wrong
When you talk to a human, this same conversation would go:
**Me:** “Hey, can you help me write a Python script to download YouTube videos?”
**Friend:** “Sure, I’ll work on that.”
**Me:** “what time is it?”
**Friend:** “3:42.”
**Me:** “Cool, how’s that script coming?”
**Friend:** “Just finished it! Here you go.”
See the difference? **The human doesn’t constantly recap what they’re doing.** They just… remember stuff and work in the background.
-----
## Current AI Problems in Voice Interactions
### Problem 1: The Awkward Recap
**Every response mentions previous context:**
- “Regarding the script we were discussing…”
- “Going back to your YouTube downloader…”
- “As I was saying about the Python code…”
**This kills conversational flow.** Humans don’t talk like this.
### Problem 2: No True Multitasking
**AI can’t work on something while talking about something else:**
- You ask for a script → AI stops everything to explain what it’s doing
- You ask about weather → AI stops again to switch topics
- You ask about the script → AI has to start over
### Problem 3: Breaking Character
**AI constantly breaks the illusion of natural conversation:**
- “I see we were working on…”
- “Getting back to your previous request…”
- “As mentioned earlier…”
**No human talks like this in casual conversation.**
-----
## My Proposed Solution
What if we split AI into two parts:
### The Big AI (Slow Thinking)
- Understands complex context
- Spawns background agents to work on tasks
- Maintains a “whiteboard” with current situation summary
### Low-Latency AI (Fast Responses)
- Handles immediate conversation
- Gets “hot-swapped” instantly for each response
- Reads the whiteboard to understand what’s happening
-----
## How This Would Work
**You:** “Write me a YouTube downloader script”
*[Big AI spawns background agent to code]*
**AI:** “I’ll create that for you.”
**You:** “What time is it?”
*[Background agent still working]*
**AI:** “3:42 PM.”
**You:** “What’s the weather like?”
*[Background agent finishes script during your question]*
**AI:** “It’s 72°F and sunny. Your YouTube script is ready!”
-----
**Alternative scenario:**
**You:** “What’s the weather like?”
*[Background agent still working]*
**AI:** “It’s 72°F and sunny.”
*[A minute later, agent finishes]*
**AI:** “Hey, your YouTube downloader script is finished!”
The AI proactively tells you when work is done - you never have to ask “are you finished yet?”
-----
## The Key Innovation: Context Compression
Instead of maintaining full conversation history, the Big AI creates an intelligent summary:
```
WHITEBOARD:
- User: Developer working on automation
- Project: YouTube downloader script
- Status: Background agent completed script
- Mood: Collaborative and focused
- Tone: Casual, technical
```
Each new low-latency instance:
1. Reads this summary instantly
1. Understands current situation
1. Responds appropriately without recap
-----
## System Instructions for Natural Behavior
Every low-latency instance gets the same behavioral rules:
```
1. Read the whiteboard to understand what's happening
2. Look at the user's question
3. If their question relates to current work → include relevant context
4. If unrelated → answer only their question
5. If background work just finished → mention it
6. Match the emotional tone
7. Never say "I see we were working on..." or similar recaps
```
-----
## Fast and Slow Thinking (Like Kahneman’s Book)
This mirrors how human cognition actually works:
- **Big AI = System 2** (slow, deliberate, analytical)
- **Low-latency AI = System 1** (fast, intuitive responses)
The Big AI does heavy thinking and compresses it into easily digestible summaries. The low-latency AI gives instant responses based on that pre-processed intelligence.
**Result: Fast responses that feel natural because they’re informed by deep understanding.**
-----
## Technical Reality
**Yes, this requires many API calls** - each response spawns a new instance.
But it could potentially run locally with efficient models. The low-latency instances don’t need to be huge - they just need to:
- Read whiteboard summaries
- Be emotionally/contextually aware
- Follow behavioral rules
- Give appropriate responses
The intelligence is in the whiteboard, not necessarily the model reading it.
-----
## What This Solves
### ✅ Natural conversation flow
**Current AI:** “Regarding the script we discussed earlier…”
**This system:** Just continues naturally
### ✅ True multitasking
**Current AI:** Stops everything to explain each step
**This system:** Works in background, mentions when done
### ✅ No awkward recaps
**Current AI:** Constantly references previous context
**This system:** Remembers without mentioning
### ✅ Emotional continuity
**Current AI:** Resets mood each response
**This system:** Maintains conversational tone
-----
## Why I Think This Matters
Right now, talking to AI feels like talking to a very smart computer that’s trying to be helpful but doesn’t understand human conversation patterns.
This approach might make AI feel like talking to a really capable person who:
- Remembers what you’re working on without constantly mentioning it
- Can handle random questions without losing track
- Works on stuff in the background and tells you when it’s done
- Just… flows naturally in conversation
**But maybe I’m wrong and people actually like the current way AI talks?— personally I HATE how current ai works and i want this to change**
| 2025-06-25T13:04:58 | https://www.reddit.com/r/LocalLLaMA/comments/1lk554m/idea_making_ai_conversations_actually_feel_like/ | anonthatisopen | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk554m | false | null | t3_1lk554m | /r/LocalLLaMA/comments/1lk554m/idea_making_ai_conversations_actually_feel_like/ | false | false | self | 0 | null |
Still early, but building a system to help AI code with full project awareness. What would help you most? | 0 | I’ve been building a tool that started out as a personal attempt to improve AI performance in programming. Over the last few weeks it’s grown a lot, and I’m planning to release a free demo soon for others to try.
The goal is to address some of the common issues that still haven’t been properly solved, things like hallucinations, lack of planning, and shallow context, especially when working on larger projects. The tool is designed for deep analysis across multi-repo or sprawling codebases where clever prompting just isn’t enough.
I’m obviously not Anthropic or OpenAI, but I think the project is starting to show real promise and I’d really like feedback from other devs who are using AI (or who gave up on it).
Specifically:
* What are the main problems you run into using LLMs for real coding?
* Can you share a time an LLM gave you a bad output, and how you fixed or worked around it?
* Any languages, frameworks, or environments where AI really struggles?
* Are there areas (like debugging, deployment, UI, profiling) where it consistently underperforms?
I’m still building, so any insight at this point would be really helpful. | 2025-06-25T13:21:59 | https://www.reddit.com/r/LocalLLaMA/comments/1lk5j69/still_early_but_building_a_system_to_help_ai_code/ | Budget_Map_3333 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk5j69 | false | null | t3_1lk5j69 | /r/LocalLLaMA/comments/1lk5j69/still_early_but_building_a_system_to_help_ai_code/ | false | false | self | 0 | null |
What are the best Speech-to-Text and Text-to-Speech models with multi lingual support? | 4 | I see a lot of SOTA models coming out, but only with English support.
What are the SOTA open source models for STT and TTS that have multilingual support ?
Is it still Whisper for speech recognition? Looking specifically por Brazilian Portuguese support to create voice agents. | 2025-06-25T13:31:47 | https://www.reddit.com/r/LocalLLaMA/comments/1lk5rl9/what_are_the_best_speechtotext_and_texttospeech/ | alew3 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk5rl9 | false | null | t3_1lk5rl9 | /r/LocalLLaMA/comments/1lk5rl9/what_are_the_best_speechtotext_and_texttospeech/ | false | false | self | 4 | null |
Nvidia DGX Spark - what's the catch? | 3 | I currently train/finetune transformer models for audio (around 50M parameters) with my mighty 3090 and for finetuning it works great, while training from scratch is close to impossible due to it being slow and not having that much VRAM.
I found out about the DGX Spark and was looking at the Asus one for $3000 but can't find what's the catch. On most places I've read about it people are complaining and saying it's not worth it and what not, but besides the slower memory bandwidth (2-3 times slower than 3090 if specs are true) - I don't see any downsides?
The most impressive thing for me is the 128GB unified memoir, which I suppose could be used as VRAM and will speed up my workflow a lot.
Is there anything to look out for when getting the DGX Spark? | 2025-06-25T13:33:53 | https://www.reddit.com/r/LocalLLaMA/comments/1lk5te5/nvidia_dgx_spark_whats_the_catch/ | lucellent | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk5te5 | false | null | t3_1lk5te5 | /r/LocalLLaMA/comments/1lk5te5/nvidia_dgx_spark_whats_the_catch/ | false | false | self | 3 | null |
The Jan.ai "team" used fake engagement to advertise their new 4B model, and deleted the post when called out | 312 | These are all of my interactions with the jan "team", followed by an instantly deleted angry comment, and the deletion of their entire announcement post without an explanation. Up to you how to interpret their response, but personally i feel i've seen enough just sorting the comment section by old and clicking a few random profiles.
[The initial interaction concerning the weirdly sycophantic comment section full of people who have apparently already used the just-announced model enough for it to be their favorite!](https://imgur.com/GHfWpXr)
[The totally real engagement from their "team members"](https://imgur.com/eds5hhX)
I'll link some of the accounts i assume are supposed to be "[team members answering people](https://imgur.com/HrtDWen)", unfortunately since [only one of the accounts pictured owned up to their affiliation](https://www.reveddit.com/v/LocalLLaMA/comments/1ljyo2p/jannano128k_a_4b_model_with_a_superlong_context/mznsoq5/#t1_mznsoq5) i'm going to have to make assumptions based on these commenters [coincidentally posting almost exclusively on jan products with posting behavior that can best be described as "suspicious".](https://imgur.com/hREcCVZ)
The original poster's account is also deleted now, which is totally what you do when you're a trustworthy org who can be trusted not to bot your github stars.
It's just especially insulting to be gaslit in this plausibly deniable way that allows them to move the goalposts as to how dishonest they were being based on how deep people are willing to dig. | 2025-06-25T13:34:38 | https://www.reddit.com/r/LocalLLaMA/comments/1lk5u1o/the_janai_team_used_fake_engagement_to_advertise/ | Xandred_the_thicc | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk5u1o | false | null | t3_1lk5u1o | /r/LocalLLaMA/comments/1lk5u1o/the_janai_team_used_fake_engagement_to_advertise/ | false | false | self | 312 | null |
Combining VRam for Inference | 1 | Given the new 5050 cards have the best vram:price ratio yet, Is it feasible to be able to combine six of them to get 48 GB of VRAM? What would the performance downsides be over 2 3090s?
Thank you! | 2025-06-25T13:34:51 | https://www.reddit.com/r/LocalLLaMA/comments/1lk5u8b/combining_vram_for_inference/ | Th3OnlyN00b | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk5u8b | false | null | t3_1lk5u8b | /r/LocalLLaMA/comments/1lk5u8b/combining_vram_for_inference/ | false | false | self | 1 | null |
Gemini CLI: your open-source AI agent | 118 | Free license gets you access to Gemini 2.5 Pro and its massive 1 million token context window. To ensure you rarely, if ever, hit a limit during this preview, we offer the industry’s largest allowance: 60 model requests per minute and 1,000 requests per day at no charge. | 2025-06-25T13:45:52 | https://blog.google/technology/developers/introducing-gemini-cli-open-source-ai-agent/ | touhidul002 | blog.google | 1970-01-01T00:00:00 | 0 | {} | 1lk63od | false | null | t3_1lk63od | /r/LocalLLaMA/comments/1lk63od/gemini_cli_your_opensource_ai_agent/ | false | false | default | 118 | {'enabled': False, 'images': [{'id': 'v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=108&crop=smart&auto=webp&s=b8e739b515523fc4b279dd605822228fe9c1b445', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=216&crop=smart&auto=webp&s=ab2e993ad8332eb7e795b22e97370eebfe347a5b', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=320&crop=smart&auto=webp&s=25bd986586ba49853f822f87350a05e8c38ed0ad', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=640&crop=smart&auto=webp&s=66dc977cf68889558dd1e0a18ef318dff22dc727', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=960&crop=smart&auto=webp&s=f55c3152e184ea246273349c731d2576601f1369', 'width': 960}, {'height': 608, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?width=1080&crop=smart&auto=webp&s=855679cab2c2648a6ee5e9a19c41da189a15ec9b', 'width': 1080}], 'source': {'height': 732, 'url': 'https://external-preview.redd.it/v_nU-59VjAFg3tUf3ktH0OR1eDLLCpt7sTIO-4lpiic.png?auto=webp&s=245c6e038d90697968891facb3165fd352f2d096', 'width': 1300}, 'variants': {}}]} |
How do I make LM Studio use the default parameters from the GGUF | 5 | I'm still quite new to the local llm space. When I look at the huggingface page of a model, there is a generation\_config.json file. This has the parameters that are loaded default onto the model, which I assume offer the best performance found by the creator.
When I download a GGUF on LM Studio I have a "Preset" loaded, I couldn't find a way to turn it off. I can create a new profile and put everything as trash but then I notice it doesn't change to the default values. I don't have any idea about the default parameters of llama.cpp (for example, the default top\_k is?) I assume that when running solely from llama.cpp it grabs the generation\_config.json from within the gguf file and automatically uses those settings + the default values if not declared.
How can I make LM Studio do the same? I have to manually go into each model page and try to see if any configuration is done, which most of the time at least the temperature is set. But then comes the issue of the rest of the parameters. Please help! | 2025-06-25T13:54:04 | https://www.reddit.com/r/LocalLLaMA/comments/1lk6axd/how_do_i_make_lm_studio_use_the_default/ | Su1tz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk6axd | false | null | t3_1lk6axd | /r/LocalLLaMA/comments/1lk6axd/how_do_i_make_lm_studio_use_the_default/ | false | false | self | 5 | null |
Looking for an upgrade from Meta-Llama-3.1-8B-Instruct-Q4_K_L.gguf, especially for letter parsing. Last time I looked into this was a very long time ago (7 months!) What are the best models nowadays? | 2 | I'm looking into LLMs for automate extracting information from letters, which are between half a page and one-and-a-half pages long most of the time. The task requires a bit of understanding and logic, but not a crazy amount.
Llama 3.1 8B does reasonably well but sometimes makes small mistakes.
I'd love to hear what similarly sized models I could use to do it slightly better.
If there are smaller, but equally good models, that'd be great, too!
I'm using llama_cpp with python bindings on a 5070ti. | 2025-06-25T13:54:23 | https://www.reddit.com/r/LocalLLaMA/comments/1lk6b7c/looking_for_an_upgrade_from/ | AuspiciousApple | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk6b7c | false | null | t3_1lk6b7c | /r/LocalLLaMA/comments/1lk6b7c/looking_for_an_upgrade_from/ | false | false | self | 2 | null |
Which gemma-3 (12b and 27b) version (Unsloth, Bartowski, stduhpf, Dampfinchen, QAT, non-QAT, etc) are you using/do you prefer? | 8 | Lately I started using different versions of Qwen-3 (I used to use the Unsloth UD ones, but recently I started moving\* to the non-UD ones or the Bartowski ones instead, as I get more t/s and more context) and I was considering the same for Gemma-3.
But between what I was reading from comments and my own tests, and I'm confused.
I remember the Bartowski, Unsloth, stduhpf, Dampfinchen, QAT, no-QAT... and reading people complaining about QAT or saying how great it is, adds to the confusion.
So, which version are you using and, if you don't mind, why? (I'm currently using the Unsloth UD ones).
\*Which I recently started to think that might be based on the different "Precision" values of the tensors, but is something I have no idea about and I still need to look at. | 2025-06-25T13:57:18 | https://www.reddit.com/r/LocalLLaMA/comments/1lk6dub/which_gemma3_12b_and_27b_version_unsloth/ | relmny | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk6dub | false | null | t3_1lk6dub | /r/LocalLLaMA/comments/1lk6dub/which_gemma3_12b_and_27b_version_unsloth/ | false | false | self | 8 | null |
What would be the best bang for my buck under $1000? | 0 | I was considering a 3090ti with 24gb of VRAM, but I would rather be steered in the right direction. Is there a better deal on the NVDA side of things?
I want to be able to set up a self hosted LLM for coding, and mess around with things like Stable Diffusion. | 2025-06-25T13:59:26 | https://www.reddit.com/r/LocalLLaMA/comments/1lk6frj/what_would_be_the_best_bang_for_my_buck_under_1000/ | Gary5Host9 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk6frj | false | null | t3_1lk6frj | /r/LocalLLaMA/comments/1lk6frj/what_would_be_the_best_bang_for_my_buck_under_1000/ | false | false | self | 0 | null |
Web search for LLMs? | 1 | Is there a way to get web search locally? | 2025-06-25T14:09:22 | https://www.reddit.com/r/LocalLLaMA/comments/1lk6p0s/web_search_for_llms/ | 00quebec | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk6p0s | false | null | t3_1lk6p0s | /r/LocalLLaMA/comments/1lk6p0s/web_search_for_llms/ | false | false | self | 1 | null |
[New Features & Better] Tabulens: A Vision-LLM Powered PDF Table Extractor | 2 | Hello everyone,
Thanks for the positive response I got on my last post about [Tabulens](https://github.com/astonishedrobo/tabulens). It really motivated me a lot to improve the package further.
>
Based on the feedback received I had already added the support for alternative model options apart from openai or google.
In the recent update:
* Previously relied on OpenCV morphology and contour analysis for table detection but now upgraded to YOLO based table detection for much higher accuracy. You can checkout the model at [https://huggingface.co/astonishedrobo/table-detection](https://huggingface.co/astonishedrobo/table-detection)
* Added and improved the validation for table extraction.
Here is the link to GitHub: [https://github.com/astonishedrobo/tabulens](https://github.com/astonishedrobo/tabulens)
You can download this as a python package.
If you test it out, I’d love any feedback or bug reports you might have. It would really help me to improve the project further.
https://preview.redd.it/5yyj22aaz29f1.png?width=640&format=png&auto=webp&s=a846b354b1754f3a803aa41f31c297f92e577f96
| 2025-06-25T14:09:26 | https://www.reddit.com/r/LocalLLaMA/comments/1lk6p2l/new_features_better_tabulens_a_visionllm_powered/ | PleasantInspection12 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk6p2l | false | null | t3_1lk6p2l | /r/LocalLLaMA/comments/1lk6p2l/new_features_better_tabulens_a_visionllm_powered/ | false | false | 2 | {'enabled': False, 'images': [{'id': '2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4.png?width=108&crop=smart&auto=webp&s=a78cacccc9c602b236cef2f648ee44d8f74fabc5', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4.png?width=216&crop=smart&auto=webp&s=852ca4da6d88848cbcfe26dc54a555972dc17185', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4.png?width=320&crop=smart&auto=webp&s=2b056fd943db47ef4d418cf3583efc81d03a6fc4', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4.png?width=640&crop=smart&auto=webp&s=4f7a1b306b0badf021b5beac72ee0e66697d6b90', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4.png?width=960&crop=smart&auto=webp&s=d4a4200fef8ab389f30896afb406d4089a9b4df6', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4.png?width=1080&crop=smart&auto=webp&s=62887a3b45785deed6419b0f24a7f0e2c7da5697', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/2rQWuZHmjTSA1RZNtS328Kw4oUVE59Uej8UFki6aih4.png?auto=webp&s=9faddb8c7c4fad168bf849a60e6b2b581522979d', 'width': 1200}, 'variants': {}}]} |
|
[Open Source] Build Your AI Team with Vibe Coding (Software 3.0 Framework) | 6 | Zentrun is an open-source Software 3.0 platform that lets you build AI agents
that grow and evolve — by creating new features through **vibe coding**.
Unlike static scripts or prompt-only tools, Zentrun agents can
**build, run, and refine** their own workflows using natural language.
From automation and analytics to full UI and database logic,
Zentrun turns your ideas into living, executable software — like real SaaS apps.
All runs locally, with full support for **MCP**, **Ollama**, and other modular backends.
---
⚡️ **Vibe-Coded AI Agents**
- Say: *“Scrape AI job posts from Reddit and send a Slack summary.”*
- Zentrun turns that into working code, stores it as a **Zent**, and lets your agent re-run or build on it.
- Each new command becomes a new skill. Your agent evolves like software — not just responds.
- Full support for local LLMs via Ollama
- Compatible with any model provider in OpenAI/Gemini/Anthropic API format
---
🧠 **Software 3.0 Architecture**
- Agents define and extend their automation, UI, analysis, and visualization — through vibe coding
- Each agent has its own embedded database — remembers state, data, and logic
- Real code execution with zero-code input: Python, browser control, API calls, shell commands
- Supports LLMs like OpenAI, Claude, Gemini, and Ollama (local)
---
🛠️ **Powered by MCP**
- Model Context Protocol handles memory, logging, and multi-tool orchestration
- Natural-language-to-execution across scraping, file parsing, DB ops, and notifications
- Zent → Agent → ZPilot hierarchy for scaling into multi-agent systems
---
💡 **Use Cases**
- Sales: auto-scrape leads, summarize contacts, send follow-ups
- HR: filter resumes, score candidates, auto-schedule interviews
- Analytics: extract → analyze → visualize — entirely with vibe-coded agents
- Marketing: generate content, monitor competitors, auto-publish across platforms
---
🖥️ **Cross-Platform, Offline, and Open Source**
- macOS, Windows, and Linux support
- Offline-first — agents work locally with full transparency
- Open-source at: https://github.com/andrewsky-labs/zentrun
---
🔗 **Explore More**
→ Try prebuilt agents or build your own AI team: https://zentrun.com
→ GitHub: https://github.com/andrewsky-labs/zentrun
---
We’re building Zentrun in public — feedback and contributions welcome!
If you’ve ever wanted an AI that grows like real software, give vibe coding a try. | 2025-06-25T15:03:34 | https://v.redd.it/3glwsm3x839f1 | mpthouse | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lk82qj | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/3glwsm3x839f1/DASHPlaylist.mpd?a=1753455828%2CZjdlNDU2OTEwNjVlZDlhMTZmMDc4ZGMzOThmMTdmNjJmY2NkYzdkNjI4NzlhZDM3YzkzMTg0NDA1YmQ0OTk4MA%3D%3D&v=1&f=sd', 'duration': 102, 'fallback_url': 'https://v.redd.it/3glwsm3x839f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/3glwsm3x839f1/HLSPlaylist.m3u8?a=1753455828%2CNTczZTIyOGUwMzI2MzJlMmRkYTM4NDdkNDBjNjljMGExZjgzMDRhNzE0NDMxYTY1MDQwMTI5YmFkOGRkMGIyOA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/3glwsm3x839f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1lk82qj | /r/LocalLLaMA/comments/1lk82qj/open_source_build_your_ai_team_with_vibe_coding/ | false | false | 6 | {'enabled': False, 'images': [{'id': 'aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p.png?width=108&crop=smart&format=pjpg&auto=webp&s=0bff12c7dfc3470e900990bfab3264d3cf70b745', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p.png?width=216&crop=smart&format=pjpg&auto=webp&s=7b29e4548813004c1936dddaee104ef3990ae4d4', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p.png?width=320&crop=smart&format=pjpg&auto=webp&s=6f6fe7d6387c9aa6bf59d00fdf2fc23914e576c2', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p.png?width=640&crop=smart&format=pjpg&auto=webp&s=4c478ea3493cc380fcd5e196c2e4509409264182', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p.png?width=960&crop=smart&format=pjpg&auto=webp&s=d2732d4189b146c7406382c7df3f3ac6cc2bc561', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p.png?width=1080&crop=smart&format=pjpg&auto=webp&s=a3a4a3c7031f93481ed0833fe090f154286be332', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/aXJtNGJuM3g4MzlmMUG656sa9a8x7y41qMK9KHse6G3IOvzv264vz6Sx8d-p.png?format=pjpg&auto=webp&s=c849ba786968ff50c39399da42d003171ba4fcd0', 'width': 1920}, 'variants': {}}]} |
|
P102-100 vs m40 12gb. Does 2gbs make much difference? | 0 | Basically it's the question in the title. How much of a difference does 2GB make? Does the newer p102-100 architecture make up for the 2GB less? | 2025-06-25T15:09:29 | https://www.reddit.com/r/LocalLLaMA/comments/1lk885n/p102100_vs_m40_12gb_does_2gbs_make_much_difference/ | EdwardRocks | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk885n | false | null | t3_1lk885n | /r/LocalLLaMA/comments/1lk885n/p102100_vs_m40_12gb_does_2gbs_make_much_difference/ | false | false | self | 0 | null |
I built an app that turns your photos into smart packing lists — all on your iPhone, 100% private, no APIs, no data collection! | 0 | Fullpack uses **Apple’s VisionKit** to extract items directly from your photos — making it easy to create packing lists, outfits, or inventory collections.
✅ Everything runs **entirely on‑device** — no APIs, no data collection.
✅ Your photos and data stay **completely private**.
Try it on the App Store — any feedback is hugely appreciated!
👉 [Fullpack on the App Store](https://apps.apple.com/us/app/fullpack/id6745692929)
# 🛠️ Tech Stack
* **Platform:** iOS‑only
* **Frontend:** SwiftUI
* **Backend:** Pure Swift
* **Database:** SwiftData
# 🎨 Design & Development
* Logo: Created with GPT‑4
* Marketing screens: Drafted in Figma
* All screens hand‑coded in SwiftUI
# 🌐 Site & Deployment
* Created site pages for the company
* Deployed in seconds via **AWS Amplify**
# 💻 Development Workflow
* 99% Xcode — handwritten code for a seamless flow
* Cursor AI used once for generating sample data
* AI = a tireless intern 😅 | 2025-06-25T15:12:52 | https://www.reddit.com/gallery/1lk8b9q | w-zhong | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1lk8b9q | false | null | t3_1lk8b9q | /r/LocalLLaMA/comments/1lk8b9q/i_built_an_app_that_turns_your_photos_into_smart/ | false | false | 0 | {'enabled': True, 'images': [{'id': 'aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA', 'resolutions': [{'height': 144, 'url': 'https://external-preview.redd.it/aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA.jpeg?width=108&crop=smart&auto=webp&s=97a39bb072425a0c51ec3b4361b3837aed65e7d9', 'width': 108}, {'height': 288, 'url': 'https://external-preview.redd.it/aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA.jpeg?width=216&crop=smart&auto=webp&s=90ddea2c5b7f32d96c44618024711033dc2e1a81', 'width': 216}, {'height': 426, 'url': 'https://external-preview.redd.it/aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA.jpeg?width=320&crop=smart&auto=webp&s=846d72c65bc1a686e5bce21e6623480571be3562', 'width': 320}, {'height': 853, 'url': 'https://external-preview.redd.it/aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA.jpeg?width=640&crop=smart&auto=webp&s=fec12a3f16e21fe2950f5955b4747bd2c7d25712', 'width': 640}, {'height': 1280, 'url': 'https://external-preview.redd.it/aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA.jpeg?width=960&crop=smart&auto=webp&s=e866cd319ca544ff2f4f6e0c9a4b2113d8f1eecc', 'width': 960}, {'height': 1440, 'url': 'https://external-preview.redd.it/aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA.jpeg?width=1080&crop=smart&auto=webp&s=6c7c16a28c05feb45b3a6749751449695844db48', 'width': 1080}], 'source': {'height': 4032, 'url': 'https://external-preview.redd.it/aXYi9YqH-FsnAu7Zwl9QYk7Gs2tECfza5gQzZlIC2WA.jpeg?auto=webp&s=1bc48053cb02d2fbf59a78396ff21c504ca426a9', 'width': 3024}, 'variants': {}}]} |
|
Correct ninja template for llama-3_3-nemotron-super-49b-v1-mlx in LMstudio? | 1 | Hi guys, I was trying to use the MLX version of Nvidia's Nemotron Super (based on Llama 3.3) but it seems like it was uploaded with an incorrect ninja template.
A solution has been suggested [here on HF](https://huggingface.co/mlx-community/Llama-3_3-Nemotron-Super-49B-v1-mlx-4bit/discussions/2), but to me it's still not clear how to fix the ninja template in LMstudio. Does anyone have the correct template, or can help me troubleshoot? Thanks! | 2025-06-25T15:50:10 | https://www.reddit.com/r/LocalLLaMA/comments/1lk9a3k/correct_ninja_template_for_llama3/ | SnowBoy_00 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lk9a3k | false | null | t3_1lk9a3k | /r/LocalLLaMA/comments/1lk9a3k/correct_ninja_template_for_llama3/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g.png?width=108&crop=smart&auto=webp&s=33c8b5eb54814d16ec363342c2829e1a2fcefd5b', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g.png?width=216&crop=smart&auto=webp&s=38b0b3eba6fa11bdb5bcd476c62b48fd7ddbbe0a', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g.png?width=320&crop=smart&auto=webp&s=eecc736e76bb431c3c696eabeb9e05b9e3db0a21', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g.png?width=640&crop=smart&auto=webp&s=d204c9d7cf7fd8519c4af1f5bfa515fde477533d', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g.png?width=960&crop=smart&auto=webp&s=a3defcc25d9caf4f486864a530f159e15786cb14', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g.png?width=1080&crop=smart&auto=webp&s=ae0efb5e96a7d749c8a9f92ae42eb19390b77dd3', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/GSIgRa1MAVqGYqtxmEItN9uVA7M4nZGK4ax8gd4ZZ3g.png?auto=webp&s=6ca25c1fb036340ab867b2f6af0c7f53d233a915', 'width': 1200}, 'variants': {}}]} |
So You Want to Learn LLMs? Here's the Roadmap | 1 | 2025-06-25T15:50:20 | https://ahmadosman.com/blog/learn-llms-roadmap/ | XMasterrrr | ahmadosman.com | 1970-01-01T00:00:00 | 0 | {} | 1lk9a9t | false | null | t3_1lk9a9t | /r/LocalLLaMA/comments/1lk9a9t/so_you_want_to_learn_llms_heres_the_roadmap/ | false | false | default | 1 | null |
|
Cydonia 24B v3.1 - Just another RP tune (with some thinking!) | 91 | Serious Note: This was really scheduled to be released today... Such awkward timing!
This official release incorporated Magistral weights through merging. It is able to think thanks to that. [Cydonia 24B v3k](https://huggingface.co/BeaverAI/Cydonia-24B-v3k-GGUF) is a proper Magistral tune but not thoroughly tested.
\---
No claims of superb performance. No fake engagements of any sort (At least I hope not. Please feel free to delete comments / downvote the post if you think it's artificially inflated). No weird sycophants.
Just a moistened up Mistral 24B 3.1, a little dumb but quite fun and easy to use! Finetuned to *hopefully* specialize on one single task: Your Enjoyment.
Enjoy! | 2025-06-25T15:59:23 | https://huggingface.co/TheDrummer/Cydonia-24B-v3.1 | TheLocalDrummer | huggingface.co | 1970-01-01T00:00:00 | 0 | {} | 1lk9ime | false | null | t3_1lk9ime | /r/LocalLLaMA/comments/1lk9ime/cydonia_24b_v31_just_another_rp_tune_with_some/ | false | false | default | 91 | {'enabled': False, 'images': [{'id': 'is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I.png?width=108&crop=smart&auto=webp&s=79e9b136777991f4ebcfdce027d019de0302e307', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I.png?width=216&crop=smart&auto=webp&s=6aba70f4acb3998d3042db52950d2699609eeee3', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I.png?width=320&crop=smart&auto=webp&s=89887946614b5f227a6c2d17730931d41d279c05', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I.png?width=640&crop=smart&auto=webp&s=8faa3dfb676a1ca98bc1f6a3369598de92ef4fba', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I.png?width=960&crop=smart&auto=webp&s=99378075bd10b2dd6d6b7f32d20682f1d6326cc8', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I.png?width=1080&crop=smart&auto=webp&s=c245fff942b69d9056ad6dbf6c99f7c59e89c1a4', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/is5dxEtYQGcop66xpu9863OAeD17dNWUu8NQ03Wo_4I.png?auto=webp&s=1b887d740853e8ef1dacb6b49adc1d128fa663df', 'width': 1200}, 'variants': {}}]} |
🚀 Revamped My Dungeon AI GUI Project – Now with a Clean Interface & Better Usability! | 19 | https://i.redd.it/20q3drcnr39f1.gif
Hey folks!
I just gave my old project [Dungeo\_ai](https://github.com/Laszlobeer/Dungeo_ai) a serious upgrade and wanted to share the improved version:
🔗 [**Dungeo\_ai\_GUI on GitHub**](https://github.com/Laszlobeer/Dungeo_ai_GUI)
This is a **local, GUI-based Dungeon Master AI** designed to let you roleplay solo DnD-style adventures using your own LLM (like a local LLaMA model via Ollama). The original project was CLI-based and clunky, but now it’s been reworked with:
🧠 **Improvements**:
* 🖥️ **User-friendly GU**I using `tkinte`r
* 🎮 More immersive roleplay support
* 💾 Easy save/load system for sessions
* 🛠️ Cleaner codebase and better modularity for community mods
* 🧩 Simple integration with local LLM APIs (e.g. Ollama, LM Studio)
🧪 Currently testing with local models like **LLaMA 3 8B/13**B, and performance is smooth even on mid-range hardware.
If you’re into solo RPGs, interactive storytelling, or just want to tinker with AI-powered DMs, I’d love your feedback or contributions!
**Try it, break it, or fork it:**
👉 [https://github.com/Laszlobeer/Dungeo\_ai\_GUI](https://github.com/Laszlobeer/Dungeo_ai_GUI)
Happy dungeon delving! 🐉 | 2025-06-25T16:48:13 | https://www.reddit.com/r/LocalLLaMA/comments/1lkau4z/revamped_my_dungeon_ai_gui_project_now_with_a/ | Reasonable_Brief578 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lkau4z | false | null | t3_1lkau4z | /r/LocalLLaMA/comments/1lkau4z/revamped_my_dungeon_ai_gui_project_now_with_a/ | false | false | 19 | {'enabled': False, 'images': [{'id': 'NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8.png?width=108&crop=smart&auto=webp&s=778bfe76d5ce65360e9679be4c52a28d3c3b1f80', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8.png?width=216&crop=smart&auto=webp&s=fbdee49226f082e85a8faef9dd39ec48f4074512', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8.png?width=320&crop=smart&auto=webp&s=7e14b60085c2626d467df6c0b0343b9d24373abc', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8.png?width=640&crop=smart&auto=webp&s=429f3f424c02aa8038222f2c0d7e58d883b099aa', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8.png?width=960&crop=smart&auto=webp&s=d4aa66d85d9d34ef9dad9d54154a6c7514f22129', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8.png?width=1080&crop=smart&auto=webp&s=cd5665eb950954832a2216105fca5f4dc4aed3b6', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/NpzXevyc8hccQld5KSdg19C9gOjRKnxTSGT0NaYuRD8.png?auto=webp&s=4b3e3ab1befd0673b932cd2d3087d05603d31551', 'width': 1200}, 'variants': {}}]} |
|
I cant see MCP in JanAI | 6 | Title, using the latest version of v0.6.1. What am i doing wrong | 2025-06-25T16:54:40 | droned-s2k | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lkb09z | false | null | t3_1lkb09z | /r/LocalLLaMA/comments/1lkb09z/i_cant_see_mcp_in_janai/ | false | false | default | 6 | {'enabled': True, 'images': [{'id': 'jmjhfoqns39f1', 'resolutions': [{'height': 60, 'url': 'https://preview.redd.it/jmjhfoqns39f1.png?width=108&crop=smart&auto=webp&s=963bb77623214613a204d78eefae5b185c3c97ea', 'width': 108}, {'height': 121, 'url': 'https://preview.redd.it/jmjhfoqns39f1.png?width=216&crop=smart&auto=webp&s=5271d5a8cc77cf9a827e4a206b4641b72cf884ff', 'width': 216}, {'height': 180, 'url': 'https://preview.redd.it/jmjhfoqns39f1.png?width=320&crop=smart&auto=webp&s=9533a6f94e50290dbb357062ab6f8de5d0eaea65', 'width': 320}, {'height': 360, 'url': 'https://preview.redd.it/jmjhfoqns39f1.png?width=640&crop=smart&auto=webp&s=9fd2937fef182a9227bcc7c2e949aa44b52d90d9', 'width': 640}, {'height': 540, 'url': 'https://preview.redd.it/jmjhfoqns39f1.png?width=960&crop=smart&auto=webp&s=38906cb31ce49e0bc76152b28d2081c87af75e7f', 'width': 960}, {'height': 607, 'url': 'https://preview.redd.it/jmjhfoqns39f1.png?width=1080&crop=smart&auto=webp&s=8994c6c360761725802a0c9d9516c51ae5f4db26', 'width': 1080}], 'source': {'height': 925, 'url': 'https://preview.redd.it/jmjhfoqns39f1.png?auto=webp&s=686cbdd62442b40ddbc7590af961e304b8b408c0', 'width': 1644}, 'variants': {}}]} |
|
Day 3 of 50 Days of Building a Small Language Model from Scratch: Building Our First Tokenizer from Scratch | 30 | 2025-06-25T16:55:08 | https://www.reddit.com/r/LocalLLaMA/comments/1lkb0r2/day_3_of_50_days_of_building_a_small_language/ | Prashant-Lakhera | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lkb0r2 | false | null | t3_1lkb0r2 | /r/LocalLLaMA/comments/1lkb0r2/day_3_of_50_days_of_building_a_small_language/ | false | false | 30 | null |
||
Llama 3.2 abliterated uncensored | 0 | Guys I'm new to artificial intelligence and I liked playing dnd style adventure games on llama 3.2 on WhatsApp before it got updated to llama 4 after which there is so much censorship, even minor gore will render the story unplayable and the ai refusestoj judge, , I tried running an abliterated uncensored version of llama 3.2 7b on my laptop using kobold with 16 gigs of ram, ryzen 5600h and an rtx 3050 4gbvram, the problem is I'm getting only 3-4 words per second and the program runs too slow despite all the adjustments I do.
Is my system too weak for it? I've tried other lighterweight forms of ai too but they face the same problem. Do I have to upgrade my system to run it properly or is there an alternative.
Please help. | 2025-06-25T16:55:46 | https://www.reddit.com/r/LocalLLaMA/comments/1lkb1ee/llama_32_abliterated_uncensored/ | DaringDebonair | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lkb1ee | false | null | t3_1lkb1ee | /r/LocalLLaMA/comments/1lkb1ee/llama_32_abliterated_uncensored/ | false | false | self | 0 | null |
Is there a better local video AI than Wan 2.1 for my 3080 12GB? No filter, of course. | 0 | Pls help | 2025-06-25T17:08:44 | https://www.reddit.com/r/LocalLLaMA/comments/1lkbdvi/is_there_a_better_local_video_ai_than_wan_21_for/ | Lower_Collection_521 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lkbdvi | false | null | t3_1lkbdvi | /r/LocalLLaMA/comments/1lkbdvi/is_there_a_better_local_video_ai_than_wan_21_for/ | false | false | self | 0 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.