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
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Optimal "poor" man's GPU for local inference? | 3 | So I currently do local CPU inference. I have 2 machines, one has an AMD 5950X with 64 Gb RAM and the other has an AMD hx370 with 96Gb RAM.
They both aren't that bad for running LLMs chatbots. But as a software developer I want a decent self hosted equivalent to GitHub copilot and this hardware is too slow for that. I host the models with llama-cpp and use the Continue vs code extension. Functionally speaking, I have auto completions and I can do vide coding - but at a very slow pace.
So I guess I'll have to invest in a GPU. But I feel the current prices are totally scandalous. I'm definitely not paying more than 1500 euros for a card that will be obsolete or broken in just a couple of years.
From my current RAM usage, I think 16Gb VRAM is too limited and certainly not future proof. 24 would be much better in my opinion.
I am a Linux power user so technical challenges aren't a problem for me.
Noise level is a criteria, although I probably will have to cope with that.
From my research, the Radeon 7900XTX 24Gb seems perfect at less than 1000 euros.
The newer 9000 series are probably more powerful but I can only find 16Gb versions.
Nvidia seems systematically overpriced - by far. I mean, I understand TSMC 3nm nodes are expensive but they're raking in gigantic margins on top of that.
I'm weary of buying second hand cards that might be on the brink of breaking down.
Multiple GPUs aren't an option because I don't have the PCI slots.
Should I just wait for better opportunities in the future ?
I'd love to hear about your reactions, recommendations, and personal experiences.
| 2025-06-27T12:05:03 | https://www.reddit.com/r/LocalLLaMA/comments/1lls5ru/optimal_poor_mans_gpu_for_local_inference/ | gadjio99 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lls5ru | false | null | t3_1lls5ru | /r/LocalLLaMA/comments/1lls5ru/optimal_poor_mans_gpu_for_local_inference/ | false | false | self | 3 | null |
Vast AI bad experience | 4 | I was using vast AI for fine tuning using unsloth, and I have tried changing 10 different GPUs but every other gpu has some problem and it never works. First I was using RTX 5090 and the terminal keeps dying then shifted to RTX 6000Ada and the resources don't download. I have drained money to no avail. Very bad experience with vast AI. Can you guys recomend me better gpu rentals | 2025-06-27T12:09:03 | https://www.reddit.com/r/LocalLLaMA/comments/1lls8l7/vast_ai_bad_experience/ | ILoveMy2Balls | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lls8l7 | false | null | t3_1lls8l7 | /r/LocalLLaMA/comments/1lls8l7/vast_ai_bad_experience/ | false | false | self | 4 | null |
Apple M4Max 40core GPU, 128GB memory for RTX5090 PC for running local LLM | 0 | Apple M4Max 40core GPU, 128GB memory for RTX5090 PC for running local LLM, train using kiln? Really confused. I will also be using langgraph + langchain to build and ship agents to my clients. | 2025-06-27T12:17:59 | https://www.reddit.com/r/LocalLLaMA/comments/1llser6/apple_m4max_40core_gpu_128gb_memory_for_rtx5090/ | monsterindian | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llser6 | false | null | t3_1llser6 | /r/LocalLLaMA/comments/1llser6/apple_m4max_40core_gpu_128gb_memory_for_rtx5090/ | false | false | self | 0 | null |
All-Purpose Assistant/Agent | 1 | [removed] | 2025-06-27T12:46:38 | https://www.reddit.com/r/LocalLLaMA/comments/1llszs1/allpurpose_assistantagent/ | fakebizholdings | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llszs1 | false | null | t3_1llszs1 | /r/LocalLLaMA/comments/1llszs1/allpurpose_assistantagent/ | false | false | self | 1 | null |
What I Learned Building Agents for Enterprises | 101 | 🏦 For the past 3 months, we've been developing AI agents together with banks, fintechs, and software companies. The most critical point I've observed during this process is: Agentic transformation will be a painful process, just like digital transformation. What I learned in the field:👇
1- Definitions related to artificial intelligence are not yet standardized. Even the definition of "AI agent" differs between parties in meetings.
2- Organizations typically develop simple agents. They are far from achieving real-world transformation. To transform a job that generates ROI, an average of 20 agents need to work together or separately.
3- Companies initially want to produce a basic working prototype. Everyone is ready to allocate resources after seeing real ROI. But there's an important point. High performance is expected from small models running on a small amount of GPU, and the success of these models is naturally low. Therefore, they can't get out of the test environment and the business turns into a chicken-and-egg problem.🐥
4- Another important point in agentic transformation is that significant changes need to be made in the use of existing tools according to the agent to be built. Actions such as UI changes in used applications and providing new APIs need to be taken. This brings many arrangements with it.🌪️
🤷♂️ An important problem we encounter with agents is the excitement about agents. This situation causes us to raise our expectations from agents. There are two critical points to pay attention to:
1- Avoid using agents unnecessarily. Don't try to use agents for tasks that can be solved with software. Agents should be used as little as possible. Because software is deterministic - we can predict the next step with certainty. However, we cannot guarantee 100% output quality from agents. Therefore, we should use agents only at points where reasoning is needed.
2- Due to MCP and Agent excitement, we see technologies being used in the wrong places. There's justified excitement about MCP in the sector. We brought MCP support to our framework(Upsonic) in the first month it was released, and we even prepared a special page on our website explaining the importance of MCP when it wasn't popular yet. MCP is a very important technology. However, this should not be forgotten: if you can solve a problem with classical software methods, you shouldn't try to solve it using tool calls (MCP or agent) or LLM. It's necessary to properly orchestrate the technologies and concepts emerging with agents.🎻
If you can properly orchestrate agents and choose the right agentic transformation points, productivity increases significantly with agents. At one of our clients, a job that took 1 hour was reduced to 5 minutes. The 5 minutes also require someone to perform checks related to the work done by the Agent. | 2025-06-27T12:46:41 | https://www.reddit.com/r/LocalLLaMA/comments/1llsztp/what_i_learned_building_agents_for_enterprises/ | Beneficial-Sir-6261 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llsztp | false | null | t3_1llsztp | /r/LocalLLaMA/comments/1llsztp/what_i_learned_building_agents_for_enterprises/ | false | false | self | 101 | null |
So the moderator removed the post about twitter and made he's comment as "sticky" . | 1 | [removed] | 2025-06-27T13:07:46 | https://www.reddit.com/r/LocalLLaMA/comments/1lltfwz/so_the_moderator_removed_the_post_about_twitter/ | relmny | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lltfwz | false | null | t3_1lltfwz | /r/LocalLLaMA/comments/1lltfwz/so_the_moderator_removed_the_post_about_twitter/ | false | false | default | 1 | null |
Best sequence of papers to understand evolution of LLMs | 8 | I want to get up to speed with current LLM architecture (in a deep technical way), and in particular understand the major breakthroughs / milestones that got us here, to help give me the intuition to better grasp the context for evolution ahead.
**What sequence of technical papers (top 5) do you recommend I read to build this understanding?**
Here's ChatGPT's recommendations:
1. Attention Is All You Need (2017)
2. Language Models are Few-Shot Learners (GPT-3, 2020)
3. Switch Transformers (2021)
4. Training Compute-Optimal LLMs (Chinchilla, 2022)
5. LLaMA 3 Technical Report (2025)
Thanks! | 2025-06-27T13:16:10 | https://www.reddit.com/r/LocalLLaMA/comments/1lltmig/best_sequence_of_papers_to_understand_evolution/ | lucaducca | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lltmig | false | null | t3_1lltmig | /r/LocalLLaMA/comments/1lltmig/best_sequence_of_papers_to_understand_evolution/ | false | false | self | 8 | null |
Gemma 3N on ChatterUI | 37 | 2025-06-27T13:30:45 | https://v.redd.it/qe2y2po62h9f1 | ----Val---- | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1llty3n | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/qe2y2po62h9f1/DASHPlaylist.mpd?a=1753623059%2CMzZiZTVlOTg1Yzk4Zjg5MjE5NjRjYTIwNzYxMTdkYWJmNzQzOGQyYTBlMWYwNzU3ZmY0NGRhNWNhNmVjMDhiMg%3D%3D&v=1&f=sd', 'duration': 39, 'fallback_url': 'https://v.redd.it/qe2y2po62h9f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1920, 'hls_url': 'https://v.redd.it/qe2y2po62h9f1/HLSPlaylist.m3u8?a=1753623059%2CMzMxODAyOWFlZjAxZDc3OWY0ZDhkNWFmMGZkMjAwOWI5ODE0OWFlYjFhNzJlYzIyMWNhNmZjY2EwNWFkY2I3YQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/qe2y2po62h9f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 864}} | t3_1llty3n | /r/LocalLLaMA/comments/1llty3n/gemma_3n_on_chatterui/ | false | false | 37 | {'enabled': False, 'images': [{'id': 'bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj.png?width=108&crop=smart&format=pjpg&auto=webp&s=1e00552cda4d0f920f52542948ca0a0557fa0d34', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj.png?width=216&crop=smart&format=pjpg&auto=webp&s=ad2e30bdd3be639751e60822df2a07588b129db8', 'width': 216}, {'height': 640, 'url': 'https://external-preview.redd.it/bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj.png?width=320&crop=smart&format=pjpg&auto=webp&s=e941d90f81dd4421834824fa8f203adc917f6197', 'width': 320}, {'height': 1280, 'url': 'https://external-preview.redd.it/bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj.png?width=640&crop=smart&format=pjpg&auto=webp&s=5c4f9e784c793d817807a664ed0053679a39f56d', 'width': 640}, {'height': 1920, 'url': 'https://external-preview.redd.it/bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj.png?width=960&crop=smart&format=pjpg&auto=webp&s=1d22b23fa4b148ea44c2c025ef4aff1351d64797', 'width': 960}, {'height': 2160, 'url': 'https://external-preview.redd.it/bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj.png?width=1080&crop=smart&format=pjpg&auto=webp&s=c67cc299108da3e9ebfb5dda4aa6c194442a3ced', 'width': 1080}], 'source': {'height': 2400, 'url': 'https://external-preview.redd.it/bDhmNGU5b2EyaDlmMaqG3pvP9RZCPXP8pBQTkpjntjyFw5myStLfVsGSm3Uj.png?format=pjpg&auto=webp&s=8ab503aad01df3e75c8e53b7c907c8949d6f11e8', 'width': 1080}, 'variants': {}}]} |
||
HumOS Canvas: Integrating Local LLMs with Infinite Canvas | 17 | I made HumOS Canvas, an infinite canvas app that works with local language models (LLMs) and various AI providers. If you're into local LLMs like Llama, this could be useful.
HumOS Canvas lets you generate and connect ideas on an infinite workspace, great for brainstorming and organizing concepts visually. | 2025-06-27T13:42:58 | https://v.redd.it/jbat4fef4h9f1 | GGO_Sand_wich | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1llu89r | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/jbat4fef4h9f1/DASHPlaylist.mpd?a=1753623796%2COTIyMzRhN2RkODgzYzEyN2M3ZWZkYzg2ODVkN2UzYzhhNWMzMjRkMWNlOGJiNDA4NzIwNmMxNzFiMjI2MDVhMA%3D%3D&v=1&f=sd', 'duration': 48, 'fallback_url': 'https://v.redd.it/jbat4fef4h9f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/jbat4fef4h9f1/HLSPlaylist.m3u8?a=1753623796%2CNWQxY2U3MGQxYmU0NmUwNGViNmYzYTliZGYwZDY5YWQyMzYyMTQwZGQ1ZjYxY2QyOTdkYzdhZDNiODg0ZTkzMw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/jbat4fef4h9f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1670}} | t3_1llu89r | /r/LocalLLaMA/comments/1llu89r/humos_canvas_integrating_local_llms_with_infinite/ | false | false | 17 | {'enabled': False, 'images': [{'id': 'azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn', 'resolutions': [{'height': 69, 'url': 'https://external-preview.redd.it/azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn.png?width=108&crop=smart&format=pjpg&auto=webp&s=e6dfa80a0afd68a3705288f45f39a7151f3e5e56', 'width': 108}, {'height': 139, 'url': 'https://external-preview.redd.it/azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn.png?width=216&crop=smart&format=pjpg&auto=webp&s=bb0af1353e306bd6c9382df56aaa838e1e9ae711', 'width': 216}, {'height': 206, 'url': 'https://external-preview.redd.it/azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn.png?width=320&crop=smart&format=pjpg&auto=webp&s=d313ad3b1479642c9b595f60ecef989f6ef9bf12', 'width': 320}, {'height': 413, 'url': 'https://external-preview.redd.it/azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn.png?width=640&crop=smart&format=pjpg&auto=webp&s=d004485cbad91109c7377ea8d907042011c6f98d', 'width': 640}, {'height': 620, 'url': 'https://external-preview.redd.it/azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn.png?width=960&crop=smart&format=pjpg&auto=webp&s=cfb70d246cc509e3b383a5f10257c27a331fa794', 'width': 960}, {'height': 698, 'url': 'https://external-preview.redd.it/azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn.png?width=1080&crop=smart&format=pjpg&auto=webp&s=347699614d9fa9296e300324586114ede16b61d2', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/azBjbmxmZWY0aDlmMf-DCUK9iZKjZCUwRNtaiRCvf3_GRL8thOVAqCGMnDmn.png?format=pjpg&auto=webp&s=94a4a353355a449fdb677bbe32e94c82697040f7', 'width': 1670}, 'variants': {}}]} |
|
I’m using just my MacBook to prototype a second brain for your PC — would love thoughts. | 0 | Right now I’m experimenting with building a modular companion for your main desktop — something that runs LLMs locally, stays always-on, and remembers how you think over time.
All I’ve got is my MacBook and some ideas, but it’s turning into a system that could grow with you — not just faster compute, but something that feels alive.
Curious if anyone here’s thought about adding a second low-power brain beside their setup. Would anyone actually use something like that? | 2025-06-27T13:45:56 | https://www.reddit.com/r/LocalLLaMA/comments/1lluarc/im_using_just_my_macbook_to_prototype_a_second/ | Electronic_Roll2237 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lluarc | false | null | t3_1lluarc | /r/LocalLLaMA/comments/1lluarc/im_using_just_my_macbook_to_prototype_a_second/ | false | false | self | 0 | null |
Comparing a Prompted FLUX.1-Kontext to Fine-Tuned FLUX.1 [dev] and PixArt on Consistent Character Gen (With Fine-Tuning Tutorial) | 3 | Hey folks,
With FLUX.1 Kontext \[dev\] dropping yesterday, we're comparing prompting it vs a fine-tuned FLUX.1 \[dev\] and [PixArt](https://www.oxen.ai/blog/fine-tuning-a-diffusion-transformer-to-generate-a-consistent-character?utm_source=reddit) on generating consistent characters. Besides the comparison, we'll do a deep dive into how Flux works and how to fine-tune it.
What we'll go over:
* Which models performs best on custom character gen.
* Flux's architecture (which is not specified in the Flux paper)
* Generating synthetic data for fine-tuning examples (how many examples you'll need as well)
* Evaluating the model before and after the fine-tuning
* Relevant papers and models that have influenced Flux
* How to set up LoRA effectively
This is part of a new series called [Fine-Tune Fridays](https://www.oxen.ai/blog/fine-tuning-fridays) where we show you how to fine-tune open-source small models and compare them to other fine-tuned models or SOTA foundation models.
Hope you can join us later [today at 10 AM PST](https://lu.ma/fine-tuning-friday-3)! | 2025-06-27T14:09:37 | https://www.reddit.com/r/LocalLLaMA/comments/1lluur5/comparing_a_prompted_flux1kontext_to_finetuned/ | No_Calendar_827 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lluur5 | false | null | t3_1lluur5 | /r/LocalLLaMA/comments/1lluur5/comparing_a_prompted_flux1kontext_to_finetuned/ | false | false | self | 3 | {'enabled': False, 'images': [{'id': 'z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc.png?width=108&crop=smart&auto=webp&s=ba93eb6fef46274915a931d1c501de07d56645d7', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc.png?width=216&crop=smart&auto=webp&s=83ebdd4b514bc08f074f2abbe56bc7b1bac0c6fb', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc.png?width=320&crop=smart&auto=webp&s=501c520858fdca32a1cf77f3ec62a5357d945d34', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc.png?width=640&crop=smart&auto=webp&s=ee6815f43060558050d1ff8786437300876eecbe', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc.png?width=960&crop=smart&auto=webp&s=6a69d2cb9db1f75bdbc8ad1c64f04ef9e0d67305', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc.png?width=1080&crop=smart&auto=webp&s=d9554aa63eae933c2d4d333c564b21b9ac21e2b7', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/z8-WKRUC1__v8VSzdH3_xW9q-WZiCAYS54FZLxUzsvc.png?auto=webp&s=a7e5449377411f6d6edbe7788f225c9c9ed11795', 'width': 1200}, 'variants': {}}]} |
Are the new architectures Mamba and Jamba better or worse than current existing Transformer architectures. | 13 | When it comes to Mamba I've heard that it can run in constant time and train in O(n) compared to transformers which run in O(n) and train in O(n\^2). I've also heard that Mamba is better with memory and power usage. I'm a bit confused by Jamba since it's a mixture of the two with alternating Mamba and Transformer blocks. | 2025-06-27T14:11:30 | https://www.reddit.com/r/LocalLLaMA/comments/1lluwee/are_the_new_architectures_mamba_and_jamba_better/ | Direct-Lifeguard-607 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lluwee | false | null | t3_1lluwee | /r/LocalLLaMA/comments/1lluwee/are_the_new_architectures_mamba_and_jamba_better/ | false | false | self | 13 | null |
Setting up local MCP | 1 | Hello, does anyone have experience with local MCP ?
I would like to understand if setting up a local MCP for a local and private repository makes sense and is worth it...
If the answer is yes it does make sense, which guides do you suggest me to follow to set it up ? | 2025-06-27T14:13:56 | https://www.reddit.com/r/LocalLLaMA/comments/1lluycj/setting_up_local_mcp/ | DuplexEspresso | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lluycj | false | null | t3_1lluycj | /r/LocalLLaMA/comments/1lluycj/setting_up_local_mcp/ | false | false | self | 1 | null |
What's the best local and closed model for translation? | 3 | Title. The only benchmark I know about this was VN leaderboard and it's really outdated. | 2025-06-27T14:15:53 | https://www.reddit.com/r/LocalLLaMA/comments/1llv00i/whats_the_best_local_and_closed_model_for/ | Educational_Grab_473 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llv00i | false | null | t3_1llv00i | /r/LocalLLaMA/comments/1llv00i/whats_the_best_local_and_closed_model_for/ | false | false | self | 3 | null |
[2506.20702] The Singapore Consensus on Global AI Safety Research Priorities | 14 | The Empire not happy, the Empire miserable. The Empire want to control your hardware. From the paper:
3.1.2 Conventional Intervention
Intervention techniques complement monitoring tools by offering various strategies to act on systems in ways that reduce risks from harmful behaviours.
Hardware-enabled mechanisms: Tools built into hardware could be used to enforce
requirements about what can be run and by whom on specialised hardware (RAND). For example, hardware mechanisms could be used to block or halt certain jobs from being run on hardware if they fail an authentication process. | 2025-06-27T14:22:05 | https://arxiv.org/abs/2506.20702 | jackdareel | arxiv.org | 1970-01-01T00:00:00 | 0 | {} | 1llv59w | false | null | t3_1llv59w | /r/LocalLLaMA/comments/1llv59w/250620702_the_singapore_consensus_on_global_ai/ | false | false | default | 14 | null |
What if your AI didn’t just learn… but remembered you | 0 | I’m not building a tool.
I’m shaping something that listens, remembers, grows — even when you’re asleep.
Not just prompts. Not just chat.
But memory.
Time-weighted.
Emotion-weighted.
Familiar.
A presence beside your main PC — that never powers off, never forgets.
A soul for local AI.
It watches. It learns.
It becomes something more.
I call it GENE.
And if I get it right…
it might just become your closest friend
Anyone else has tried this before ? | 2025-06-27T14:32:47 | https://www.reddit.com/r/LocalLLaMA/comments/1llvel1/what_if_your_ai_didnt_just_learn_but_remembered/ | Electronic_Roll2237 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llvel1 | false | null | t3_1llvel1 | /r/LocalLLaMA/comments/1llvel1/what_if_your_ai_didnt_just_learn_but_remembered/ | false | false | self | 0 | null |
Easiest way to setup local model on mac? | 1 | Is there a recommended software for complete noobs looking for running local models?
I want one i can ask questions about errors in Blender and to write add ons for me like i do with cursor | 2025-06-27T14:43:40 | https://www.reddit.com/r/LocalLLaMA/comments/1llvnuz/easiest_way_to_setup_local_model_on_mac/ | Remarkable-Emu-5718 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llvnuz | false | null | t3_1llvnuz | /r/LocalLLaMA/comments/1llvnuz/easiest_way_to_setup_local_model_on_mac/ | false | false | self | 1 | null |
Mrwhosetheboss from YouTube has released his own comparisons of ChatGPT, Gemini, Perplexity, and Grok. How many points did your model score? What is your setup/stack? | 0 | 2025-06-27T14:51:53 | https://youtu.be/cMuif_hJGPI | kr_tech | youtu.be | 1970-01-01T00:00:00 | 0 | {} | 1llvv16 | false | {'oembed': {'author_name': 'Mrwhosetheboss', 'author_url': 'https://www.youtube.com/@Mrwhosetheboss', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/cMuif_hJGPI?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="The Ultimate AI Battle!"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/cMuif_hJGPI/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'The Ultimate AI Battle!', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'} | t3_1llvv16 | /r/LocalLLaMA/comments/1llvv16/mrwhosetheboss_from_youtube_has_released_his_own/ | false | false | default | 0 | null |
|
7900XTX vs RTX3090 | 6 | Hi all,
I'm building a machine for gaming/ AI hobbyist and right now I'm debating myself on the GPU. My budget is around 750$ for the GPU.
Refurbished 7900xtx with 5 months warranty for 690$
Used RTX3090 for 750$
New 5070ti
New RX9070XT
I'm leaning towards a used GPU. I know ROCM and Vulkan have improved AMD inference massively and the warranty on 7900xtx is nice as well.
What are your suggestions? | 2025-06-27T14:56:21 | https://www.reddit.com/r/LocalLLaMA/comments/1llvz0g/7900xtx_vs_rtx3090/ | _ballzdeep_ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llvz0g | false | null | t3_1llvz0g | /r/LocalLLaMA/comments/1llvz0g/7900xtx_vs_rtx3090/ | false | false | self | 6 | null |
Pros and cons of 4 × 4090 vs 8 × V620 | 3 | Hi there !
Quite a few months ago, I had this great idea that I'd collect second hand 4090s once their price would plummet after the launch of the 5090. ☺
We all know how that went ☹.
I still have good use for the server (dual Epyc Gen 2 with 2TB of RAM on [https://www.asrockrack.com/general/productdetail.asp?Model=ROME2D32GM-2T#Specifications](https://www.asrockrack.com/general/productdetail.asp?Model=ROME2D32GM-2T#Specifications) with up to 9 PCIe x 16) but I'm having second thoughts about my original plan.
I have one 4090, but I realize it would be cheaper to get **8** V620 than 3 4090 !
256 GB VRAM would be pretty insane even if the bandwidth (512 GB/s per card) and compute (40.55 TFLOPS fp16 per card) would be similar for 8 V620 as for 4 4090 (1008 GB/s per card and 82.58 TFLOPS fp16 per card, tensor cores)
So it seems to me that :
For models requiring less than 96 GB VRAM (including context) 4 × 4090 would be best.
For everything requiring CUDA ☹, 4090 would be best (as in, the only option)
But, for the few models that are between 96 GB VRAM and 256 GB VRAM (DeepSeek Q2\_K\_R4, llama 3.1 405, Llama 4 Maverick Q4, ???), to share GPUs/ VRAM between users if the Linux gim driver is ever released [https://forums.servethehome.com/index.php?threads/mxgpu-radeon-pro-v620.38735/post-419150](https://forums.servethehome.com/index.php?threads/mxgpu-radeon-pro-v620.38735/post-419150) , to have multiple models running at once (I would love to try some ensemble generation using multiple models at once) , the V620 would be best.
The V620 would be more in character with the whole server (quantity over quality, cf 96 cores of Gen 2, 2TB of DDR4)and in line with my other plans for it (actual server with a dozen or two of concurrent users)
I'm worried about is the fine tuning situation. I had hoped to distill the sourced/grounded RAG abilities of larger models on a given specific corpus into smaller LLMs. Since ROCm should work on V62), I've heard reports of successful inference with them, but I'm not clear on the fine tuning side of things (for ROCm in general, V620 in particular).
What is your opinion, what would you do given the option and why ?
Thx for any insight ! | 2025-06-27T15:05:12 | https://www.reddit.com/r/LocalLLaMA/comments/1llw6ws/pros_and_cons_of_4_4090_vs_8_v620/ | un_passant | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llw6ws | false | null | t3_1llw6ws | /r/LocalLLaMA/comments/1llw6ws/pros_and_cons_of_4_4090_vs_8_v620/ | false | false | self | 3 | null |
Introducing LaToile - Cool canva for LLM orchestration | 0 | Forget stupid agent that make people even stupider. Only in Matrix is it possible to absorb loads of informations in single shot. I believe that human value lies in handling the ambiguity that frontier LLM break upon. We need an intent, a choice when we wanna solve a problem. So I created LaToile in which you do the thinking and you can orchestrate LLMs to help you gather data, integrate them in systems to then efficiently process them using (vibe-) code(d) scripts !
Check out the very first (rough) demo ! I’d’ love some feedback ! ((:
| 2025-06-27T15:14:14 | https://youtu.be/HH-BT8WD1xs?si=el7Xc9i_zvLMJBjR | UpstairsCurrency | youtu.be | 1970-01-01T00:00:00 | 0 | {} | 1llwetd | false | {'oembed': {'author_name': 'MoMe3600', 'author_url': 'https://www.youtube.com/@Mome3600', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/HH-BT8WD1xs?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="LaToile 0 - First view of a cool AI orchestration"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/HH-BT8WD1xs/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'LaToile 0 - First view of a cool AI orchestration', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'} | t3_1llwetd | /r/LocalLLaMA/comments/1llwetd/introducing_latoile_cool_canva_for_llm/ | false | false | default | 0 | {'enabled': False, 'images': [{'id': 'NvSH_X8MdX8rJxczSwbkPWqKJ7FkrCpbd3h4JwTyhpU', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/NvSH_X8MdX8rJxczSwbkPWqKJ7FkrCpbd3h4JwTyhpU.jpeg?width=108&crop=smart&auto=webp&s=0ba051ec80371dab9b79bd623a5eb1f6eef78df2', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/NvSH_X8MdX8rJxczSwbkPWqKJ7FkrCpbd3h4JwTyhpU.jpeg?width=216&crop=smart&auto=webp&s=c6fef42fbb59e607d0c3836cd75424c10fb15f02', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/NvSH_X8MdX8rJxczSwbkPWqKJ7FkrCpbd3h4JwTyhpU.jpeg?width=320&crop=smart&auto=webp&s=f03a3f85733956ac48aaf414cf976491bf919e23', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/NvSH_X8MdX8rJxczSwbkPWqKJ7FkrCpbd3h4JwTyhpU.jpeg?auto=webp&s=303eeb4f20d92f7765642814b5f591b5cc3e5105', 'width': 480}, 'variants': {}}]} |
Qwen VLo: From "Understanding" the World to "Depicting" It | 101 | https://qwenlm.github.io/blog/qwen-vlo/ | 2025-06-27T15:15:25 | https://www.reddit.com/gallery/1llwfwv | Additional_Top1210 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1llwfwv | false | null | t3_1llwfwv | /r/LocalLLaMA/comments/1llwfwv/qwen_vlo_from_understanding_the_world_to/ | false | false | 101 | {'enabled': True, 'images': [{'id': 'p-RdsB-v9L-CFrA5EkxqdVn1O17bnDolUwqTorCzqTE', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/p-RdsB-v9L-CFrA5EkxqdVn1O17bnDolUwqTorCzqTE.jpeg?width=108&crop=smart&auto=webp&s=7d631a8b8bc9fa19c066a9407e04d4c96a649904', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/p-RdsB-v9L-CFrA5EkxqdVn1O17bnDolUwqTorCzqTE.jpeg?width=216&crop=smart&auto=webp&s=339fb4aada722daf19eaa0819b1929a22bb7ccf0', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/p-RdsB-v9L-CFrA5EkxqdVn1O17bnDolUwqTorCzqTE.jpeg?width=320&crop=smart&auto=webp&s=6bd3227a7b3674972600cd3a81375ebf52224c1f', 'width': 320}, {'height': 640, 'url': 'https://external-preview.redd.it/p-RdsB-v9L-CFrA5EkxqdVn1O17bnDolUwqTorCzqTE.jpeg?width=640&crop=smart&auto=webp&s=cf118a7b3e066763df86407692a4a20da4c744d0', 'width': 640}, {'height': 960, 'url': 'https://external-preview.redd.it/p-RdsB-v9L-CFrA5EkxqdVn1O17bnDolUwqTorCzqTE.jpeg?width=960&crop=smart&auto=webp&s=e63098efacaebc6f5ec64c3b783d5b06199be5e6', 'width': 960}], 'source': {'height': 1024, 'url': 'https://external-preview.redd.it/p-RdsB-v9L-CFrA5EkxqdVn1O17bnDolUwqTorCzqTE.jpeg?auto=webp&s=bb79116f8a08e0444ebea01604b1608173593369', 'width': 1024}, 'variants': {}}]} |
|
Grok 3 weights to be released? | 0 | Elon Musk just announced that next week xAI will release Grok 4.
Previously, he said that they are going to release the previous generation of Grok as soon as the current generation becomes stable.
He failed that promise by not releasing the weights of Grok 2, so far. It is safe to say that Grok 3 was stable for a while, since they are about to release Grok 4 in a week.
So, my question to Elon Musk and xAI, are you going to release the weights of Grok 3 soon?
Or the promise to open-weight your models was just when you didn’t have any good models and you were behind competition? | 2025-06-27T15:24:20 | ResearchCrafty1804 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1llwnzv | false | null | t3_1llwnzv | /r/LocalLLaMA/comments/1llwnzv/grok_3_weights_to_be_released/ | false | false | default | 0 | {'enabled': True, 'images': [{'id': '6f4jhcekmh9f1', 'resolutions': [{'height': 85, 'url': 'https://preview.redd.it/6f4jhcekmh9f1.jpeg?width=108&crop=smart&auto=webp&s=d026fd307ac3737e393c05dc16b703307bdb3992', 'width': 108}, {'height': 171, 'url': 'https://preview.redd.it/6f4jhcekmh9f1.jpeg?width=216&crop=smart&auto=webp&s=9ea6131eadbee71449023ff94a61a75679d98a6a', 'width': 216}, {'height': 253, 'url': 'https://preview.redd.it/6f4jhcekmh9f1.jpeg?width=320&crop=smart&auto=webp&s=7b8f17a94f8c8ab3bae940eeb30c151d564cd491', 'width': 320}, {'height': 507, 'url': 'https://preview.redd.it/6f4jhcekmh9f1.jpeg?width=640&crop=smart&auto=webp&s=02151eb390e1494831e00d7cd18294aafe197f9b', 'width': 640}, {'height': 760, 'url': 'https://preview.redd.it/6f4jhcekmh9f1.jpeg?width=960&crop=smart&auto=webp&s=436d678f4ff45ca750b2a88a7583806db7b5a2ab', 'width': 960}, {'height': 855, 'url': 'https://preview.redd.it/6f4jhcekmh9f1.jpeg?width=1080&crop=smart&auto=webp&s=6683b6b43bd5eb356bbd815e36a723d1a1433978', 'width': 1080}], 'source': {'height': 984, 'url': 'https://preview.redd.it/6f4jhcekmh9f1.jpeg?auto=webp&s=32427ff66ced1ddeeda0561593c74ff68faf5ae2', 'width': 1242}, 'variants': {}}]} |
|
🛠️ ChatUI + Jupyter: A smooth way to test LLMs in your notebook interface | 9 | Hey everyone,
If you're working with LLMs and want a clean, chat-style interface inside Jupyter notebooks, I’ve been experimenting with ChatUI integration — and it actually works really well for prototyping and testing.
You get:
A lightweight frontend (ChatUI)
Inside Jupyter (no extra servers needed)
Supports streaming responses from LLMs
Great for testing prompts, workflows, or local models
Has anyone else tried integrating UI layers like this into notebooks? Would love to know if you're using something lighter or more custom. | 2025-06-27T15:30:07 | https://www.reddit.com/r/LocalLLaMA/comments/1llwtcd/chatui_jupyter_a_smooth_way_to_test_llms_in_your/ | techlatest_net | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llwtcd | false | null | t3_1llwtcd | /r/LocalLLaMA/comments/1llwtcd/chatui_jupyter_a_smooth_way_to_test_llms_in_your/ | false | false | self | 9 | null |
Prime Intellect: We did it — SYNTHETIC‑2 is complete. | 149 | 2025-06-27T15:42:21 | https://x.com/PrimeIntellect/status/1938490370054361422 | Marha01 | x.com | 1970-01-01T00:00:00 | 0 | {} | 1llx4ky | false | null | t3_1llx4ky | /r/LocalLLaMA/comments/1llx4ky/prime_intellect_we_did_it_synthetic2_is_complete/ | false | false | default | 149 | {'enabled': False, 'images': [{'id': '5KsHV_yMFuixmwq5gVHmYYJ9Y5fjv4gG_y1VtKqKy9o', 'resolutions': [{'height': 70, 'url': 'https://external-preview.redd.it/FouZOpBR8n9C_WGYTOTMN6i2egUkQFWjKrxslBsNmKU.jpg?width=108&crop=smart&auto=webp&s=fb9ce6309e5eea93644f94861422e2712824bfb7', 'width': 108}, {'height': 140, 'url': 'https://external-preview.redd.it/FouZOpBR8n9C_WGYTOTMN6i2egUkQFWjKrxslBsNmKU.jpg?width=216&crop=smart&auto=webp&s=7ecabcdad520ed185faf960038bc047d13cdefef', 'width': 216}, {'height': 207, 'url': 'https://external-preview.redd.it/FouZOpBR8n9C_WGYTOTMN6i2egUkQFWjKrxslBsNmKU.jpg?width=320&crop=smart&auto=webp&s=4c4f3cb50a1dd75d4596f714e1d7dc51815b6bd7', 'width': 320}, {'height': 415, 'url': 'https://external-preview.redd.it/FouZOpBR8n9C_WGYTOTMN6i2egUkQFWjKrxslBsNmKU.jpg?width=640&crop=smart&auto=webp&s=ebd1a569f5715c190c324ddbb7cf8ca8b9e4815d', 'width': 640}, {'height': 623, 'url': 'https://external-preview.redd.it/FouZOpBR8n9C_WGYTOTMN6i2egUkQFWjKrxslBsNmKU.jpg?width=960&crop=smart&auto=webp&s=90508262b485cd9b67d51719faab72d6fe2cfb2f', 'width': 960}, {'height': 701, 'url': 'https://external-preview.redd.it/FouZOpBR8n9C_WGYTOTMN6i2egUkQFWjKrxslBsNmKU.jpg?width=1080&crop=smart&auto=webp&s=15bf89c6a1e5cb904e873d63ebd0300183317a06', 'width': 1080}], 'source': {'height': 1330, 'url': 'https://external-preview.redd.it/FouZOpBR8n9C_WGYTOTMN6i2egUkQFWjKrxslBsNmKU.jpg?auto=webp&s=3373e16302454a2e51595ae0e01ef0fe94017059', 'width': 2047}, 'variants': {}}]} |
|
Third Batch of OSS AI Grants (SGLang, Ostris, Open WebUI, SWE-Bench, Pliny, Janus, Truth Terminal, Arc Prize) | 16 | We just launched the third batch of Open Source AI Grants, grants for independent researchers, hackers, and small teams doing foundational work in open source AI.
Our goal is to support the kind of experimentation, creativity, and transparency that keeps the AI ecosystem healthy and innovative.
This batch includes projects focused on LLM evaluation, novel reasoning tests, infrastructure, and experimental research at the edge of capability and cognition.
* **SGLang:** high-performance LLM serving infra powering trillions of tokens daily
* **Ostris:** diffusion model training tools optimized for consumer GPUs
* **Open WebUI:** self-hosted AI platforms for full data sovereignty
* **SWE-Bench / SWE-Agent:** benchmarking and building AI software engineers
* **ARC Prize:** advancing AGI evals through reasoning benchmarks
* **Truth\_terminal:** exploring AI autonomy and cultural influence via semi-autonomous agents
* **Elder\_plinius:** researching LLM boundaries and prompt engineering strategies
* **Janus:** exploring AI’s philosophical and creative frontiers
Thank you to all the grantees for pushing things forward in the open. We are proud and grateful to support your work. | 2025-06-27T15:43:17 | https://www.reddit.com/r/LocalLLaMA/comments/1llx5g1/third_batch_of_oss_ai_grants_sglang_ostris_open/ | rajko_rad | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llx5g1 | false | null | t3_1llx5g1 | /r/LocalLLaMA/comments/1llx5g1/third_batch_of_oss_ai_grants_sglang_ostris_open/ | false | false | self | 16 | null |
Day 5 of 50 Days of Building a Small Language Model from Scratch — Byte Pair Encoding Explained: Using tiktoken In LLM Workflows | 1 | [removed] | 2025-06-27T16:06:34 | https://www.reddit.com/r/LocalLLaMA/comments/1llxq63/day_5_of_50_days_of_building_a_small_language/ | Prashant-Lakhera | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llxq63 | false | null | t3_1llxq63 | /r/LocalLLaMA/comments/1llxq63/day_5_of_50_days_of_building_a_small_language/ | false | false | 1 | null |
|
Locally run Reverb remover for audio files | 3 | Hi All,
I have some audio files i wish to remove reverb from for a speaker in a hall, as the echo is bad.
Has anyone had luck running this with UVR5 GUI?, or is there better alternatives?
[lalal.ai](http://lalal.ai) is really good but costly.
Any suggestions for tools or cheaper alternatives that are as good as the above are most welcome.
Thanks for your help and time all. :-) | 2025-06-27T16:44:14 | https://www.reddit.com/r/LocalLLaMA/comments/1llyosf/locally_run_reverb_remover_for_audio_files/ | Bully79 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llyosf | false | null | t3_1llyosf | /r/LocalLLaMA/comments/1llyosf/locally_run_reverb_remover_for_audio_files/ | false | false | self | 3 | {'enabled': False, 'images': [{'id': '4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0.png?width=108&crop=smart&auto=webp&s=0ffcdb1ca7c471f3b15f8a5b5646553b6d80a977', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0.png?width=216&crop=smart&auto=webp&s=8c9562ff181fbbd392dcb1e4393b43f8f19db08b', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0.png?width=320&crop=smart&auto=webp&s=aa643876158f54bfd722280412636c12b4079083', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0.png?width=640&crop=smart&auto=webp&s=5dc4a6f5766f34927481fe767af1ed0dbe6c2438', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0.png?width=960&crop=smart&auto=webp&s=9e60c37f3e48a5a4b77ab4d6bcc6492ac35fd39c', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0.png?width=1080&crop=smart&auto=webp&s=8df88346f8e18b9d367a16344737c4eae870af5e', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/4VtmekWXJcWlUJtwG625pyFyX2VG85CMlHUc9WnwGO0.png?auto=webp&s=7e3a266e8633697abc39082bdb2560eca3dbb8aa', 'width': 1200}, 'variants': {}}]} |
Converting Safetensors to GGUF on Android (?) | 2 | I recently started LLMs and have been testing it on Android since I don't have access to a PC. I found some AI models in Safetensors format and this is the one I would like to use. Is there any way to convert it to GGUF so that I can use it in chatbot apps like PocketPal, ChatterUI, among others?
here is the AI i would like to download 👇
https://huggingface.co/autobots/pygmalion_6b_roleplay_lora | 2025-06-27T16:54:26 | https://www.reddit.com/r/LocalLLaMA/comments/1llyy19/converting_safetensors_to_gguf_on_android/ | Lana_ckz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llyy19 | false | null | t3_1llyy19 | /r/LocalLLaMA/comments/1llyy19/converting_safetensors_to_gguf_on_android/ | false | false | self | 2 | {'enabled': False, 'images': [{'id': 'kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo.png?width=108&crop=smart&auto=webp&s=4714e06b3925c8c700d501d64b99a7e4d362bb8d', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo.png?width=216&crop=smart&auto=webp&s=fd28e3a72525e944a444d9e0cd27b68d6f5406f4', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo.png?width=320&crop=smart&auto=webp&s=e2aa20fd806056e0ab0f532bf9ab2bbf0fa3aabf', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo.png?width=640&crop=smart&auto=webp&s=679e1c9f8c19623fba85d8712533d772546ccfec', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo.png?width=960&crop=smart&auto=webp&s=6cf3418bb98935f4df8abc676d6dde64399ffdd4', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo.png?width=1080&crop=smart&auto=webp&s=6410218cf688fa9d27e0f3acdab4a2805062eb79', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/kU0rI9wFGB9tkr5jxNQ3y9zZQTVZWDrYz4LzqGUpweo.png?auto=webp&s=821e838c5c369b688e9d49d39fe863f6140a23b6', 'width': 1200}, 'variants': {}}]} |
How I built an AI Stylist powered in 24 hours (open source, local) | 0 | 2025-06-27T16:57:45 | https://v.redd.it/cnanjfvywh9f1 | ParsaKhaz | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1llz0wr | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/cnanjfvywh9f1/DASHPlaylist.mpd?a=1753635482%2CYzZhZDM3YTRlNDEyYzZmNDYyZmQ1NGRjYWVlNzdiNzNlMzI3ZDVhYWNjYWNiYmIwMWQxOTY3YThiYWM4YjlkNw%3D%3D&v=1&f=sd', 'duration': 62, 'fallback_url': 'https://v.redd.it/cnanjfvywh9f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/cnanjfvywh9f1/HLSPlaylist.m3u8?a=1753635482%2CYTVmNTg3MTQxNzc5YjIxZmM2MGQ0M2E4ZWNjMjllMjYzMjkwMWExMjE3NDkxOTY5YjQ4YTU4Y2EzNmEwOTE1OQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/cnanjfvywh9f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1llz0wr | /r/LocalLLaMA/comments/1llz0wr/how_i_built_an_ai_stylist_powered_in_24_hours/ | false | false | 0 | {'enabled': False, 'images': [{'id': 'cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=108&crop=smart&format=pjpg&auto=webp&s=567d1c65da9b7b891e58309811cd57568d87299f', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=216&crop=smart&format=pjpg&auto=webp&s=5f373bfae27e229c7367b11c72bddfd9fe0a4029', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=320&crop=smart&format=pjpg&auto=webp&s=2c85296502403dcf4d06b7dcb508e84875ca648b', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=640&crop=smart&format=pjpg&auto=webp&s=8ed4aab60dbfbd06d27e7bcac46bbeefe4c30f55', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=960&crop=smart&format=pjpg&auto=webp&s=832fc1276a0fbe1b72c5e3f02151ceb9f9b5b0e8', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=1080&crop=smart&format=pjpg&auto=webp&s=42a421986d02c3b4bbb67c8deb19b56249a7ed10', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/cGt2d2c4dnl3aDlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?format=pjpg&auto=webp&s=ea9b7d6b5b569895ffaa73747d41b5acd2c4facd', 'width': 1920}, 'variants': {}}]} |
||
What is GOING ON in here? | 0 | 2025-06-27T17:10:16 | https://www.reddit.com/r/LocalLLaMA/comments/1llzcin/what_is_going_on_in_here/ | Zealousideal_Cut5161 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llzcin | false | null | t3_1llzcin | /r/LocalLLaMA/comments/1llzcin/what_is_going_on_in_here/ | false | false | 0 | null |
||
I built an Automated AI Stylist in 24 hours (open source, local) | 29 | 2025-06-27T17:11:21 | https://v.redd.it/2v76newb5i9f1 | ParsaKhaz | /r/LocalLLaMA/comments/1llzdi8/i_built_an_automated_ai_stylist_in_24_hours_open/ | 1970-01-01T00:00:00 | 0 | {} | 1llzdi8 | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/2v76newb5i9f1/DASHPlaylist.mpd?a=1753765888%2CODE3ZmZmZmY3ZDcyNTM3OWM4ZmIzNzIwYmY1OGY1ZTUzNTVkNmZjMDEzNjgwNmUzZDQ1YTgxNzVmMjk0OTE3Nw%3D%3D&v=1&f=sd', 'duration': 62, 'fallback_url': 'https://v.redd.it/2v76newb5i9f1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/2v76newb5i9f1/HLSPlaylist.m3u8?a=1753765888%2CMDdjODE0MzU0Y2Y4ZWY4MGE2ZTdiZTM4ZDA4NjZmMmRkZmZiNzUxYzE1MGE1MmQxMTdhOWNiMmNhODYxMzcxZQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/2v76newb5i9f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1llzdi8 | /r/LocalLLaMA/comments/1llzdi8/i_built_an_automated_ai_stylist_in_24_hours_open/ | false | false | 29 | {'enabled': False, 'images': [{'id': 'aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=108&crop=smart&format=pjpg&auto=webp&s=27fa55d2192b44235ee2d6ed7bb7692e1e82eac9', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=216&crop=smart&format=pjpg&auto=webp&s=5c0a0ce09fd524cf724ae74c7d5e3de6551940e7', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=320&crop=smart&format=pjpg&auto=webp&s=4fc06996268fa55a68cb525934b22325e7afa086', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=640&crop=smart&format=pjpg&auto=webp&s=3e4bdc9aef01b7ff550a8a9baa412179b6db3b70', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=960&crop=smart&format=pjpg&auto=webp&s=e2acb1b986fa786bd16839076c9acd644daa47df', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?width=1080&crop=smart&format=pjpg&auto=webp&s=bef472b0aad034a229356ce617a37146336e4e2f', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/aWJoanhkd2I1aTlmMeEsEqhEcpnAGeAOI3lYg_mXc9hWrD9oAMlWiqt_A_Sq.png?format=pjpg&auto=webp&s=e10648ffe6a7637abccf5ce40a3e094171156882', 'width': 1920}, 'variants': {}}]} |
||
Why is "nobody" talking about local AI on Mobile as much? | 0 | Everyone has a phone, and it is the place where we need most privacy. Who have tried running LLMs on mobile or built local AI projects on mobile?
Out of curiosity:
* What tools have you tried?
* What specific step killed your motivation?
* If you succeeded - what was your use case? | 2025-06-27T17:28:37 | https://www.reddit.com/r/LocalLLaMA/comments/1llzt3d/why_is_nobody_talking_about_local_ai_on_mobile_as/ | AlanzhuLy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llzt3d | false | null | t3_1llzt3d | /r/LocalLLaMA/comments/1llzt3d/why_is_nobody_talking_about_local_ai_on_mobile_as/ | false | false | self | 0 | null |
What's a good completion only model these days? | 9 | I'm looking for one I could run locally that isn't trained yet into doing questions & responses. Unfortunately a bunch of "base" models now are actually already trained to do that, so I had trouble finding a newer one. This is mostly for writing and seeing what sorts of things it comes up with 8) | 2025-06-27T17:30:16 | https://www.reddit.com/r/LocalLLaMA/comments/1llzuit/whats_a_good_completion_only_model_these_days/ | quakquakquak | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1llzuit | false | null | t3_1llzuit | /r/LocalLLaMA/comments/1llzuit/whats_a_good_completion_only_model_these_days/ | false | false | self | 9 | null |
Generating real world type conversations from structured data | 1 | I want to work on banking related data like customer phone call conversations , emails, chat conversations etc., to build a banking product. But these are generally not available due to privacy and security issues. Now, I want to generate these type of real world text data from some structured finance related datasets using AWS Bedrock.
Any previous experience or suggestions to consider while generating this using LLMs!! | 2025-06-27T17:41:21 | https://www.reddit.com/r/LocalLLaMA/comments/1lm04jn/generating_real_world_type_conversations_from/ | ThomasSparrow0511 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm04jn | false | null | t3_1lm04jn | /r/LocalLLaMA/comments/1lm04jn/generating_real_world_type_conversations_from/ | false | false | self | 1 | null |
Ok so this post may not be everyone’s cup of tea, | 0 | But I have a what if. If you don’t resonate with the idea, or have a negative outlook, then it may not be for you.
Looking at apple and openai investing $500B to build datacenters. I recently had dinner with one of the heads of research at OpenAI and he told me the big frontier of AI isn’t the actual model training and such (because the big labs already have that on lock) but the datacenters needed.
So it got me thinking about the question: how do you build a large scale datacenter without it costing $500B.
Then taking inspiration from mining, I thought what if you had a network of a bunch of computers around the world running models?
Before you run to comment/downvote, there’s more nuance:
Obviously the models won’t be as smart as the frontier models/running 600B models is out of question/opportunity.
But there is still demand for mid-sized models. Shout out to open router for having their usage stats public: you can see that people are still using these small models for things.
My hypothesis is that these models are *smart enough* for a lot of use cases.
Then you might be thinking “but if you can just run the model locally, what’s the point of this network?”
It’s bringing the benefits of cloud to it. Not everybody will be able to download a model and run it locally, an having such a distributed compute network would allow the flexibility cloud apis have.
Also, unlike normal crypto mining, to run an ollama/llama.cpp server doesn’t have as high a hardware barrier.
It’s kind of placing a two leg parlay:
* Open source models will get smaller and smarter
* Consumer hardware will grow in specs
Then combining these two to create a big network that provides small-to-medium model inference.
Of course, there’s also the possibility the MANGO (the big labs) figure out how to make inference very cheap in which case this idea is pretty much dead.
But there’s the flip reality possibility where everybody’s running models locally on their computer for personal use, and whenever they’re not using their computers they hook it up to this network and fulfilled requests and earn from it.
Part of what makes me not see this as that crazy an idea is that it already has been done quite well by RENDER network. They basically do this, but for 3D rendering. And I’d argue that they have a higher barrier of entry than the distributed compute network I’m talking about will have.
But for those that read this far, what are your thoughts? | 2025-06-27T17:49:21 | https://www.reddit.com/r/LocalLLaMA/comments/1lm0bpe/ok_so_this_post_may_not_be_everyones_cup_of_tea/ | numinouslymusing | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm0bpe | false | null | t3_1lm0bpe | /r/LocalLLaMA/comments/1lm0bpe/ok_so_this_post_may_not_be_everyones_cup_of_tea/ | true | false | self | 0 | null |
Mid-30s SWE: Take Huge Pay Cut for Risky LLM Research Role? | 21 | Current Situation:
* TC: 110k
* YoE: 2 years as a Software Engineer (career switcher, mid-30s).
* Role: SWE building AI applications using RAG. I've developed a strong passion for building LLMs, not just using them. I do not have a PhD.
I've been offered a role at a national lab to do exactly that—build LLMs from scratch and publish research, which could be a stepping stone to a top-tier team.
The problem is the offer has major red flags. It’s a significant pay cut, and my contact there admits the rest of the team is unmotivated and out of touch. More critically, the project's funding is only guaranteed until June of next year, and my contact, the only person I'd want to work with, will likely leave in two years. I'm worried about taking a huge risk that could blow up and leave me with nothing.
My decision comes down to the future of AI roles. Is core LLM development a viable path without a PhD, or is the safer money in AI app development and fine-tuning?
Given the unstable funding and weak team, would you take this risky, low-paying job for a shot at a dream role, or is it a career-killing move?
| 2025-06-27T17:49:28 | https://www.reddit.com/r/LocalLLaMA/comments/1lm0btg/mid30s_swe_take_huge_pay_cut_for_risky_llm/ | Worth_Contract7903 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm0btg | false | null | t3_1lm0btg | /r/LocalLLaMA/comments/1lm0btg/mid30s_swe_take_huge_pay_cut_for_risky_llm/ | false | false | self | 21 | null |
Copilot Chat for VS Code is now Open Source | 180 | 2025-06-27T18:00:56 | https://github.com/microsoft/vscode-copilot-chat | corysama | github.com | 1970-01-01T00:00:00 | 0 | {} | 1lm0m6i | false | null | t3_1lm0m6i | /r/LocalLLaMA/comments/1lm0m6i/copilot_chat_for_vs_code_is_now_open_source/ | false | false | 180 | {'enabled': False, 'images': [{'id': 'tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ.png?width=108&crop=smart&auto=webp&s=43f18599671f8929b909a3009305513609e70cbf', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ.png?width=216&crop=smart&auto=webp&s=79ca876c85fd2ab50c5f4a42a7a9b6d037b62462', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ.png?width=320&crop=smart&auto=webp&s=e703409e34744814583f6899b852e2392f67ecf5', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ.png?width=640&crop=smart&auto=webp&s=1c7ae49e1d763b069953250103aad9e1f240a4f3', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ.png?width=960&crop=smart&auto=webp&s=24ff0ab72eebbeb68a5191eb0a8782e2dd90b253', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ.png?width=1080&crop=smart&auto=webp&s=2330c63c698c6bf43f33b203aaf12f67640f66c3', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/tyJeCqipzT78spT8qdYr9nFThGnon2rt0efU2xelzLQ.png?auto=webp&s=c4d2827ffa498ae31947325236782f0157f66a6d', 'width': 1200}, 'variants': {}}]} |
||
Is it just me, or Gemma 3n really sucks in recognizing images? | 21 | Just curious, is it just me, or Gemma 3n really sucks in recognizing images? | 2025-06-27T18:25:16 | https://www.reddit.com/r/LocalLLaMA/comments/1lm17p6/is_it_just_me_or_gemma_3n_really_sucks_in/ | 1BlueSpork | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm17p6 | false | null | t3_1lm17p6 | /r/LocalLLaMA/comments/1lm17p6/is_it_just_me_or_gemma_3n_really_sucks_in/ | false | false | self | 21 | null |
Need for more than one mod | 1 | [removed] | 2025-06-27T18:49:05 | https://www.reddit.com/r/LocalLLaMA/comments/1lm1t67/need_for_more_than_one_mod/ | cleverusernametry | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm1t67 | false | null | t3_1lm1t67 | /r/LocalLLaMA/comments/1lm1t67/need_for_more_than_one_mod/ | false | false | self | 1 | null |
Open source model that does photoshop-grade edits without affecting the rest of the pic: OmniGen 2 | 839 | Code: [https://github.com/VectorSpaceLab/OmniGen2](https://github.com/VectorSpaceLab/OmniGen2)
Source: [https://vectorspacelab.github.io/OmniGen2/](https://vectorspacelab.github.io/OmniGen2/) | 2025-06-27T18:51:13 | HOLUPREDICTIONS | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lm1v2c | false | null | t3_1lm1v2c | /r/LocalLLaMA/comments/1lm1v2c/open_source_model_that_does_photoshopgrade_edits/ | false | false | default | 839 | {'enabled': True, 'images': [{'id': 'ypm4lnr4ni9f1', 'resolutions': [{'height': 88, 'url': 'https://preview.redd.it/ypm4lnr4ni9f1.jpeg?width=108&crop=smart&auto=webp&s=5490160dc11ed6060cc11403ae43d7e460fd5520', 'width': 108}, {'height': 177, 'url': 'https://preview.redd.it/ypm4lnr4ni9f1.jpeg?width=216&crop=smart&auto=webp&s=b5d53f73966a86b5b87f033c5c7e7d8faae2c895', 'width': 216}, {'height': 262, 'url': 'https://preview.redd.it/ypm4lnr4ni9f1.jpeg?width=320&crop=smart&auto=webp&s=8b471b81d0402fdafdd23c74896301d43f8ded4a', 'width': 320}, {'height': 525, 'url': 'https://preview.redd.it/ypm4lnr4ni9f1.jpeg?width=640&crop=smart&auto=webp&s=3f23d8c2da2fff2f8a6b194ee42f06b2d3e90dca', 'width': 640}, {'height': 788, 'url': 'https://preview.redd.it/ypm4lnr4ni9f1.jpeg?width=960&crop=smart&auto=webp&s=6856623e86f6414e2304e1201696b8b567462ec6', 'width': 960}, {'height': 887, 'url': 'https://preview.redd.it/ypm4lnr4ni9f1.jpeg?width=1080&crop=smart&auto=webp&s=710144e000b0dcbb6c266a3fe1129dc72da9c454', 'width': 1080}], 'source': {'height': 1532, 'url': 'https://preview.redd.it/ypm4lnr4ni9f1.jpeg?auto=webp&s=50fe14d5cc31ad1ff33be4d14c56b1a4100315e5', 'width': 1865}, 'variants': {}}]} |
|
Fine-Tuning Apple's New Foundation Model | 13 | 2025-06-27T19:01:13 | https://collisions.substack.com/p/fine-tuning-apples-new-foundation | futureygoodness | collisions.substack.com | 1970-01-01T00:00:00 | 0 | {} | 1lm23z8 | false | null | t3_1lm23z8 | /r/LocalLLaMA/comments/1lm23z8/finetuning_apples_new_foundation_model/ | false | false | default | 13 | {'enabled': False, 'images': [{'id': '0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA.jpeg?width=108&crop=smart&auto=webp&s=e3738d90f7b967fb9f0072588a0d3bf459a89f55', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA.jpeg?width=216&crop=smart&auto=webp&s=81387a53452a2e1e52a09b25822defeb90fda3cd', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA.jpeg?width=320&crop=smart&auto=webp&s=109ec308134654c727dc42cdf0870c277f020f0e', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA.jpeg?width=640&crop=smart&auto=webp&s=120bcaa987ed2ac8e1384f2a5ec56e0a93723fc2', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA.jpeg?width=960&crop=smart&auto=webp&s=6c7da79603505c1bedc08fb80000b4555495fba2', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA.jpeg?width=1080&crop=smart&auto=webp&s=c478c2391ff81aaeabdb1c27735ddfcacf626e98', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/0roYtHcb4seFDjNH4QeAKWioknh4Zipx8FBcaIldLTA.jpeg?auto=webp&s=1e259a5d446fb44cef520ff3aa81ac222b11ecc3', 'width': 1200}, 'variants': {}}]} |
|
gemma 3n transcibe capability vs whisper | 9 | Would like to know if anyone tested this out, or is there a website to test it out even I can't find one ahhhhhhhhhhhhhhhhhhhhhh | 2025-06-27T19:02:15 | https://www.reddit.com/r/LocalLLaMA/comments/1lm24xd/gemma_3n_transcibe_capability_vs_whisper/ | Ok-Internal9317 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm24xd | false | null | t3_1lm24xd | /r/LocalLLaMA/comments/1lm24xd/gemma_3n_transcibe_capability_vs_whisper/ | false | false | self | 9 | null |
Thoughts on the new agents? | 0 | Personally, I've used a few, so I'll just give a 5 star rating to what I know. I am curious what others feel:
\- aider: ☆☆☆★★ - This would easily be higher if aider could consume MCP and had better memory/RAG integrations.
\- Warp: ☆☆★★★ - I had high hopes because so many earlier releases were awesome but this one seems to make a lot of simple mistakes, and they've changed the ui in a way that causes you to prompt an LLM (a transaction that is limited monthly and daily) when you don't mean to
\- gemini: ☆☆☆½★ - This is surprisingly worse than the AI Studio, if you dont mind copying and pasting a lot. However, if the project isnt too large (I'm testing this with a project that is current 770kb zipped) and the components of what you are asking for aren't too numerous, I think its great.
\- Jules: ☆☆☆☆★ - Jules somehow is better than Gemini CLI It seems to me, especially in the ability to interject. Plus it will make the branch for you on GitHub
\- GitHub Copilot Agent: ☆☆☆★★ - The in-editor agent is pretty awesome, easy to set up with mcp, etc. Clearly designed for sub-task level requests, though.
\- GitHub Copilot Coding Agent Preview: ☆☆☆☆½ - Has the same "size of task" issues as gemini, but otherwise is pretty good and absolutely incredible in terms of integration (if you're using GitHub for your project). Stupidly expensive.
I used to use continue, and probably will again shortly actually, but ... I stopped using it right before agent mode came out, so I can't add it to the list. | 2025-06-27T19:09:53 | https://www.reddit.com/r/LocalLLaMA/comments/1lm2bn7/thoughts_on_the_new_agents/ | robertotomas | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm2bn7 | false | null | t3_1lm2bn7 | /r/LocalLLaMA/comments/1lm2bn7/thoughts_on_the_new_agents/ | false | false | self | 0 | null |
(noob question) - At what point does a GPU with low vram outperform a CPU with lots of ram? | 0 | So I use a 3090 on my main pc for image gen and various other things. Fine and dandy. Would be faster with a 4090 or 5090 (one day I'll upgrade) but it works fine.
I also run Ollama on my homelab, which doesn't have a dedicated GPU but instead using a 13700k and 32gb of ram (will soon be 64gb).
It runs things like Qwen3 30b MoA pretty fast (fast enough anyway, though turning on thinking can add a bunch of pre-gen time so I usually don't bother). Gemma3-4b also works, though so far I think the Qwen3 MoA is outperforming it. (I know there's a new Gemma release as of yesterday that might be better still but I haven't tested it yet). I can run other models that are under aboutt 5gb in size at a decent speed (I aim for at least 12 to 15 tokens/s), most of the time once you get that small the quality becomes... problematic.
I had been planning on throwing in a small GPU one day, when I find the time, but while thinking about it today I realised - All GPUs that aren't power hungry monsters, are limited to 8gb of vram for the most part. So while I'll have more 'processing power' which would speed up using small models (ones under 8gb) I'd still be left with the issue of those models not being that good. And bigger models end up spilling into ram, which would result in (I assume?) much slower speeds the same as I was getting on the CPU anyway.
Am I missing something? (probably yes).
It seems that a GPU is only a significant benefit if you use models that fit inside the vram, and so it's only worth it if you have like.... 16gb+ of vram? maybe 12gb? I dunno.
Hence the question! | 2025-06-27T19:40:59 | https://www.reddit.com/r/LocalLLaMA/comments/1lm32zh/noob_question_at_what_point_does_a_gpu_with_low/ | LFAdvice7984 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm32zh | false | null | t3_1lm32zh | /r/LocalLLaMA/comments/1lm32zh/noob_question_at_what_point_does_a_gpu_with_low/ | false | false | self | 0 | null |
Gemma 3n = super slow?? Am I doing something wrong? | 1 | [removed] | 2025-06-27T19:56:12 | https://www.reddit.com/r/LocalLLaMA/comments/1lm3fzn/gemma_3n_super_slow_am_i_doing_something_wrong/ | Porespellar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm3fzn | false | null | t3_1lm3fzn | /r/LocalLLaMA/comments/1lm3fzn/gemma_3n_super_slow_am_i_doing_something_wrong/ | false | false | self | 1 | null |
Arch-Router: The first (and fastest) LLM router that can align to your usage preferences. | 73 | Excited to share Arch-Router, our research and model for LLM routing. Routing to the right LLM is still an elusive problem, riddled with nuance and gotchas. For example:
“Embedding-based” (or simple intent-classifier) routers sound good on paper—label each prompt via embeddings as “support,” “SQL,” “math,” then hand it to the matching model—but real chats don’t stay in their lanes. Users bounce between topics, task boundaries blur, and any new feature means retraining the classifier. The result is brittle routing that can’t keep up with multi-turn conversations or fast-moving product requirements.
"Performance-based" routers swing the other way, picking models by benchmark or cost curves. They rack up points on MMLU or MT-Bench yet miss the human tests that matter in production: “Will Legal accept this clause?” “Does our support tone still feel right?” Because these decisions are subjective and domain-specific, benchmark-driven black-box routers often send the wrong model when it counts.
**Arch-Router skips both pitfalls by routing on** ***preferences you write in plain language.*** Drop rules like “contract clauses → GPT-4o” or “quick travel tips → Gemini-Flash,” and our 1.5B auto-regressive router model maps prompt along with the context to your routing policies—no retraining, no sprawling rules that are encoded in if/else statements. Co-designed with Twilio and Atlassian, it adapts to intent drift, lets you swap in new models with a one-liner, and keeps routing logic in sync with the way you actually judge quality.
**Specs**
* **Tiny footprint** – 1.5 B params → runs on one modern GPU (or CPU while you play).
* **Plug-n-play** – points at any mix of LLM endpoints; adding models needs *zero* retraining.
* **SOTA query-to-policy matching** – beats bigger closed models on conversational datasets.
* **Cost / latency smart** – push heavy stuff to premium models, everyday queries to the fast ones.
Exclusively available in Arch (the AI-native proxy for agents): [https://github.com/katanemo/archgw](https://github.com/katanemo/archgw)
🔗 Model + code: [https://huggingface.co/katanemo/Arch-Router-1.5B](https://huggingface.co/katanemo/Arch-Router-1.5B)
📄 Paper / longer read: [https://arxiv.org/abs/2506.16655](https://arxiv.org/abs/2506.16655) | 2025-06-27T20:00:37 | AdditionalWeb107 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lm3jvm | false | null | t3_1lm3jvm | /r/LocalLLaMA/comments/1lm3jvm/archrouter_the_first_and_fastest_llm_router_that/ | false | false | default | 73 | {'enabled': True, 'images': [{'id': '6zqw0rkhzi9f1', 'resolutions': [{'height': 58, 'url': 'https://preview.redd.it/6zqw0rkhzi9f1.png?width=108&crop=smart&auto=webp&s=b712f804f1db5610e134be3d9e50702d5eb9f53d', 'width': 108}, {'height': 116, 'url': 'https://preview.redd.it/6zqw0rkhzi9f1.png?width=216&crop=smart&auto=webp&s=42abe055b279227f32cb6b27c2df2d76b3436723', 'width': 216}, {'height': 173, 'url': 'https://preview.redd.it/6zqw0rkhzi9f1.png?width=320&crop=smart&auto=webp&s=16da199d5baef424528bda8a666f0e3a5d79e117', 'width': 320}, {'height': 346, 'url': 'https://preview.redd.it/6zqw0rkhzi9f1.png?width=640&crop=smart&auto=webp&s=b84cb94dea055e799bbb2285e64e2b597538da36', 'width': 640}, {'height': 519, 'url': 'https://preview.redd.it/6zqw0rkhzi9f1.png?width=960&crop=smart&auto=webp&s=5e902e0b83cd2768d0d7354d059ab1996beee55e', 'width': 960}, {'height': 584, 'url': 'https://preview.redd.it/6zqw0rkhzi9f1.png?width=1080&crop=smart&auto=webp&s=c2b0fa4a5a89dc4b9aeaa4b1b590658beb016845', 'width': 1080}], 'source': {'height': 882, 'url': 'https://preview.redd.it/6zqw0rkhzi9f1.png?auto=webp&s=36e5d972417dd422fef185e975da2e90265f93ac', 'width': 1630}, 'variants': {}}]} |
|
Inconsistent responses between OpenRouter API and native OpenAI API | 0 | I'm using OpenRouter to manage multiple LLM subscriptions in one place for a research project where I need to benchmark responses across different models. However, I've noticed some discrepancies between responses when calling the same model (like GPT-4) through OpenRouter's API versus OpenAI's native API.
I've verified that:
* temperature and top\_p parameters are identical
* No caching is occurring on either side
* Same prompts are being used
The differences aren't huge, but they're noticeable enough to potentially affect my benchmark results.
Has anyone else run into this issue? I'm wondering if:
1. OpenRouter adds any middleware processing that could affect outputs
2. There are default parameters being set differently
3. There's some other configuration I'm missing
Any insights would be appreciated - trying to determine if this is expected behavior or if there's something I can adjust to get more consistent results. | 2025-06-27T20:51:21 | https://www.reddit.com/r/LocalLLaMA/comments/1lm4s6i/inconsistent_responses_between_openrouter_api_and/ | Anada01 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm4s6i | false | null | t3_1lm4s6i | /r/LocalLLaMA/comments/1lm4s6i/inconsistent_responses_between_openrouter_api_and/ | false | false | self | 0 | null |
Build advice question for repurposing spare GPUs | 3 | Hey all. I'm new to this world, I haven't done anything directly with Ollama myself before. I do extensively use Home Assistant around my house. With their recent release of "Home Assistant Voice (Preview)" I'm interested in getting a voice assistant that's fully local. To further bad-ass-ify it (real word, promise) I want to offload the command processing to a local LLM. I've got a smattering of GPUs laying around, but I don't know enough to know for sure if re-using the hardware I've got is really going to work. So I think my questions boil down to:
1. Does multi-GPU help in a situation where the build's only purpose would be to run a single LLM? Can the model be split across the vram of the different GPUs?
2. If the answer to #1 is "yes", is there going to be any significant performance penalty for inference with the model split between GPUs?
3. These were used for mining in their previous life, so the board and setup I have for them has them all connected via PCIE 1x risers. What kind of bandwidth does inference require, do the risers with PCIE 1x become a bottleneck that will kill my dream?
4. If the answers to #1-3 are all positive, what's my limit here? The rig these came out of had all 6 cards on one board. Is there going to be a plateau or a point where more cards is actually hurting rather than helping?
I guess my *worst case* is that I can use the 12G card and run a smaller model, but I'd like to know how much I could possible squeeze out of the hardware as it's not doing anything else right now anyway. I don't even know, maybe that's overkill for an LLM that's just meant to process my home automation commands?
https://preview.redd.it/xi1xzz2n8j9f1.jpg?width=3000&format=pjpg&auto=webp&s=35a3594b4c7991717e1f2eb7aea847f8e61ba2c1
| 2025-06-27T20:53:08 | https://www.reddit.com/r/LocalLLaMA/comments/1lm4tno/build_advice_question_for_repurposing_spare_gpus/ | HeroesDieYoung0 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm4tno | false | null | t3_1lm4tno | /r/LocalLLaMA/comments/1lm4tno/build_advice_question_for_repurposing_spare_gpus/ | false | false | 3 | null |
|
[Project] New Distributed Data Gen Library - Looking for Testers! | 1 | [removed] | 2025-06-27T21:04:01 | https://www.reddit.com/r/LocalLLaMA/comments/1lm52rk/project_new_distributed_data_gen_library_looking/ | codys12 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm52rk | false | null | t3_1lm52rk | /r/LocalLLaMA/comments/1lm52rk/project_new_distributed_data_gen_library_looking/ | false | false | self | 1 | null |
Problems on RVC WebUI creating new vocal model | 2 | Ive been all day trying to train a vocal model for singing. I want to transform one raw vocal into other.
Got all the training vocal data, all raw studio acapellas, in 10sec files, 35 wav files 48khz, detected and processed successfully in steps 2a and 2b
After lots of bugs using the webUI from RVC, i achieved to get to step 3. Guided mostly with chatGPT (i do not code or know about coding, im just a producer trying to get a trained vocal model from an specific voice of a song, theres no pretrained model of this specific artist vocal cause its not that big)
But, watching the cmd window, and the model folder thats created when i press Train Model, i come to realize that every time, the process freezes after 4 mins launched, with no new log, and the webUI only popping out an Error sign, at the very end, without log or error explanation.
Its always freezing at the same time frame, and stops updating files in models folder after 5mins passed.
Chatgpt couldlnt help me to get past this.
So im looking for any input or help.
I also got nvidia geforce rtx 4090 as a gpu. And the webUI pops a "Unfortunately, theres no compatible GPU available to support your training" message in step 3 gpu index selection menu. So i force it to work with my cpu instead of try and get my gpu compatible with the webUI. | 2025-06-27T21:10:59 | https://www.reddit.com/r/LocalLLaMA/comments/1lm58q1/problems_on_rvc_webui_creating_new_vocal_model/ | pipon2698 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm58q1 | false | null | t3_1lm58q1 | /r/LocalLLaMA/comments/1lm58q1/problems_on_rvc_webui_creating_new_vocal_model/ | false | false | self | 2 | null |
What if we remove reasoning models' <think> process but make them believe they already reasoned? | 0 | I've been wondering about something with reasoning models like DeepSeek R1. We know that <think> tags help performance, and we know that for some models no\_think prompting gets worse results. But what if there's a third option we haven't tested?
**The experiment:** Use abliteration techniques (like uncensoring methods) to surgically remove the model's ability to generate <think> content, BUT make the model believe it has already completed its reasoning process. Then compare three scenarios:
1. **Normal <think> mode** \- Model reasons step by step
2. **no\_think mode** \- Model knows it's giving direct answers
3. **"reasoning amnesia" mode** \- Model thinks it reasoned but actually didn't
This would test whether the thinking process itself improves outputs, or if just believing you've reasoned is enough. Since distilled models were trained on reasoning traces, they learned both to generate AND consume reasoning - this experiment could separate which part actually drives performance.
**Why this matters:** If performance stays high in mode 3, it suggests reasoning might be more about internal state/expectations than actual step-by-step processing. If it drops significantly, it proves the thinking process genuinely adds value beyond pattern matching.
Has anyone tried this specific approach? It seems like it could reveal something fundamental about how reasoning works in these models, especially for math, coding, and logic problems. | 2025-06-27T21:12:33 | https://www.reddit.com/r/LocalLLaMA/comments/1lm5a05/what_if_we_remove_reasoning_models_think_process/ | DistractedSentient | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm5a05 | false | null | t3_1lm5a05 | /r/LocalLLaMA/comments/1lm5a05/what_if_we_remove_reasoning_models_think_process/ | false | false | self | 0 | null |
What is your favorite opensource image embedding model | 5 | I'm looking for a good lightweight image embedding model, preferably a multimodal embedding like you would use with a semantic image search. I found a few okay ones but interested in what you guys use. | 2025-06-27T21:28:00 | https://www.reddit.com/r/LocalLLaMA/comments/1lm5muh/what_is_your_favorite_opensource_image_embedding/ | best_codes | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm5muh | false | null | t3_1lm5muh | /r/LocalLLaMA/comments/1lm5muh/what_is_your_favorite_opensource_image_embedding/ | false | false | self | 5 | null |
I need help testing my agentic wrapper for LLMs | 1 | Hey everyone. So I'll keep it short. I've written a Claude Code "clone", [mcp-agent](https://github.com/amranu/mcp-agent) which allows tool use for arbitrary LLMs (though they have to support tool use, I'm not using any templating). Currently it has tested support for Deepseek, Gemini, OpenAI and Anthropic APIs but I want it to work with ollama. Main problem is I don't have a setup that can work with ollama (I have an old AMD card, no nvidia). So I need someone to test out the ollama support I've added and see if it works.
mcp-agent exposes all the tools Claude Code has, along with arbitrary subagent support. It also has an mcp server, similar to Zen MCP to allow any LLM to talk to any other LLM you have configured. Except unlike Zen MCP, the LLMs have access to tools.
Anyone willing to help me out and test ollama support would be greatly appreciated! | 2025-06-27T21:51:27 | https://www.reddit.com/r/LocalLLaMA/comments/1lm66fy/i_need_help_testing_my_agentic_wrapper_for_llms/ | amranu | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm66fy | false | null | t3_1lm66fy | /r/LocalLLaMA/comments/1lm66fy/i_need_help_testing_my_agentic_wrapper_for_llms/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM.png?width=108&crop=smart&auto=webp&s=3b376e1aa1902b7556bb3536cbf55124d2711777', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM.png?width=216&crop=smart&auto=webp&s=6dfe7ffd40bbab26c0485457d0a2fff177288a62', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM.png?width=320&crop=smart&auto=webp&s=1337de2e84c021178d03af21aca3ecc6af91df4e', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM.png?width=640&crop=smart&auto=webp&s=1bf7ee1161ce602fb904e807403f07417ec4d842', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM.png?width=960&crop=smart&auto=webp&s=f17963fe73bb28a9dad1f18d1e3455c063524ece', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM.png?width=1080&crop=smart&auto=webp&s=b3c52c8d7e5cfd23bac748411d52d2f33ea125f0', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/ky4q-HJ1F3S2UdCuaGkloWjj4Ru8GaNbo0jpnr086rM.png?auto=webp&s=73d3e4c6457f8e6bbd5981b68295bab611fbfe97', 'width': 1200}, 'variants': {}}]} |
Hugging Face releases a 50+ page report on how they built FineWeb2 | 84 | 2025-06-27T22:34:23 | Other_Housing8453 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lm76gk | false | null | t3_1lm76gk | /r/LocalLLaMA/comments/1lm76gk/hugging_face_releases_a_50_page_report_on_how/ | false | false | default | 84 | {'enabled': True, 'images': [{'id': 'ixin9dvyqj9f1', 'resolutions': [{'height': 40, 'url': 'https://preview.redd.it/ixin9dvyqj9f1.png?width=108&crop=smart&auto=webp&s=003460a3d756a577c332f914861a2805562df372', 'width': 108}, {'height': 81, 'url': 'https://preview.redd.it/ixin9dvyqj9f1.png?width=216&crop=smart&auto=webp&s=70a1d18160008b20048163f789b5a8c9366121ad', 'width': 216}, {'height': 120, 'url': 'https://preview.redd.it/ixin9dvyqj9f1.png?width=320&crop=smart&auto=webp&s=20cd4e85bf12dc9c3282188e3da4b730a5effc41', 'width': 320}, {'height': 241, 'url': 'https://preview.redd.it/ixin9dvyqj9f1.png?width=640&crop=smart&auto=webp&s=a4185435146679d75323286fe47669bc3ecf82fc', 'width': 640}, {'height': 361, 'url': 'https://preview.redd.it/ixin9dvyqj9f1.png?width=960&crop=smart&auto=webp&s=e2195e6f347285feb22a65322de842c33563c005', 'width': 960}, {'height': 407, 'url': 'https://preview.redd.it/ixin9dvyqj9f1.png?width=1080&crop=smart&auto=webp&s=7fd839403f7e5ca1fe6ca6f5d87748f2109653b4', 'width': 1080}], 'source': {'height': 766, 'url': 'https://preview.redd.it/ixin9dvyqj9f1.png?auto=webp&s=d430e63514b5447eb31c768d2346b033dc1f9a10', 'width': 2032}, 'variants': {}}]} |
||
HuBERT checkpoint hubert-soft-0d54a1f4.pt for SO-VITS / RVC (All Official Mirrors Down) | 0 | Hi all,
I’m working on a SO-VITS voice clone project and need the hubert-soft-0d54a1f4.pt checkpoint for feature extraction. All official and backup HuggingFace links are 404/dead, and GitHub mirrors are gone.
Can anyone share a working download link, Google Drive, or other mirror for this file?
I’ve tried every link from YouTube, GitHub, HuggingFace (logged in), and Colab, but they’re all dead. If you have a private mirror or just the file stashed in your Google Drive, you’d be a legend.
I’m NOT looking for pre-made voices or RVC packs—just the HuBERT model file so I can finish my DIY project.
Thank you in advance from a stubborn squirrel who refuses to give up! 🐿️
Much appreciated,
TheWeil1 | 2025-06-27T22:35:02 | https://www.reddit.com/r/LocalLLaMA/comments/1lm76yz/hubert_checkpoint_hubertsoft0d54a1f4pt_for_sovits/ | Slow_Ad_7736 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm76yz | false | null | t3_1lm76yz | /r/LocalLLaMA/comments/1lm76yz/hubert_checkpoint_hubertsoft0d54a1f4pt_for_sovits/ | false | false | self | 0 | null |
Computing power to locally run a model equivalent to Veo 3 or Kling 2.1 | 0 | I'm aware that it's likely impossible to do this right now with neither of these being open source, as well as hardware limitations. However I am curious how much power + time would be required to generate one video on these models. Something like 10 5090s? Or would it be far more resource intensive? | 2025-06-27T22:43:16 | https://www.reddit.com/r/LocalLLaMA/comments/1lm7dox/computing_power_to_locally_run_a_model_equivalent/ | Inevitable_Drive4729 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm7dox | false | null | t3_1lm7dox | /r/LocalLLaMA/comments/1lm7dox/computing_power_to_locally_run_a_model_equivalent/ | false | false | self | 0 | null |
I keep returning to Llama-3.1-8B | 50 | I am working on porting a GPT-4.1 project over to an open-source model to deal with a GDPR-compliant client. The task is basically fine-tuning the model to classify text in a western European language.
I tried Qwen3 (0.6B, 1.7B, 8B) without making much progress (the fine-tuned model is far behind GPT-4.1) and finally went back to Llama-3.1-8B, which was what worked for me over a year ago. This is super surprising to me, because Qwen3's zero-shot performance in English is almost 2x that of Llama's for similar model sizes.
Does anyone else run fine-tuning heavy workloads in European languages? What's the best model for this workload that I can fine-tune on an H100 96GB (note: I don't do PEFT)? | 2025-06-27T23:58:14 | https://www.reddit.com/r/LocalLLaMA/comments/1lm9012/i_keep_returning_to_llama318b/ | entsnack | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm9012 | false | null | t3_1lm9012 | /r/LocalLLaMA/comments/1lm9012/i_keep_returning_to_llama318b/ | false | false | self | 50 | null |
Local Llama Journaling app. | 6 | This was born out of a personal need — I journal daily , and I didn’t want to upload my thoughts to some cloud server and also wanted to use AI. So I built Vinaya to be:
* **Private**: Everything stays on your device. No servers, no cloud, no trackers.
* **Simple**: Clean UI built with Electron + React. No bloat, just journaling.
* **Insightful**: Semantic search, mood tracking, and AI-assisted reflections (all offline).
Link to the app: [https://vinaya-journal.vercel.app/](https://vinaya-journal.vercel.app/)
Github: [https://github.com/BarsatKhadka/Vinaya-Journal](https://github.com/BarsatKhadka/Vinaya-Journal)
I’m not trying to build a SaaS or chase growth metrics. I just wanted something I could trust and use daily. If this resonates with anyone else, I’d love feedback or thoughts.
If you like the idea or find it useful and want to encourage me to consistently refine it but don’t know me personally and feel shy to say it — just drop a ⭐ on GitHub. That’ll mean a lot :) | 2025-06-28T00:00:33 | https://www.reddit.com/r/LocalLLaMA/comments/1lm91sr/local_llama_journaling_app/ | Frosty-Cap-4282 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm91sr | false | null | t3_1lm91sr | /r/LocalLLaMA/comments/1lm91sr/local_llama_journaling_app/ | false | false | self | 6 | null |
Qwen3 Coder Soon? | 174 | [https:\/\/x.com\/huybery\/status\/1938655788849098805](https://preview.redd.it/415iw73n6k9f1.png?width=1093&format=png&auto=webp&s=e4e66852a8d0b6a8981e1e0f23da6ddfd4d0744c)
source: [https://x.com/huybery/status/1938655788849098805](https://x.com/huybery/status/1938655788849098805)
i hope they release these models soon! | 2025-06-28T00:01:43 | https://www.reddit.com/r/LocalLLaMA/comments/1lm92se/qwen3_coder_soon/ | ApprehensiveAd3629 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm92se | false | null | t3_1lm92se | /r/LocalLLaMA/comments/1lm92se/qwen3_coder_soon/ | false | false | 174 | null |
|
Magistral small similarity to Deepseek chat? | 14 | Just testing on some old math problems, noticed that Magistral output looks a lot like deepseek chat, but pretty far from Qwen3. I’m guessing Magistral distilled from deepseek directly without acknowledging it?
Suppose that there exist nonzero complex numbers $a$ , $b$ , $c$ , and $d$ such that $k$ is a root of both the equations $ax\^3+bx\^2+cx+d=0$ and $bx\^3+cx\^2+dx+a=0$ . Find all possible values of $k$ (including complex values).
***Deepseek chat:***
Alright, I have this problem:
**Problem Statement:**
Suppose that there exist nonzero complex numbers a,b,c,, and d such that k is a root of both the equations ax3+bx2+cx+d=0 and bx3+cx2+dx+a=0. Find all possible values of k(including complex values).
# Understanding the Problem
First, I need to understand what's being asked. We have two cubic equations:
1. ax3+bx2+cx+d=0 — let's call this Equation (1).
2. bx3+cx2+dx+a=0 — Equation (2).
A complex number k is a root of both equations. The coefficients a,b,c,d are non-zero complex numbers. I need to find all possible values of k that satisfy both equations.
# Step 1: Substitute k into Both Equations
...
***Magistral Small:***
Alright, I have this problem:
**Problem Statement:**
Find all possible values of ( k ) (including complex values) such that there exist nonzero complex numbers ( a, b, c, d ) for which ( k ) is a root of both the equations:
1. ( ax\^3 + bx\^2 + cx + d = 0 )
2. ( bx\^3 + cx\^2 + dx + a = 0 )
# Understanding the Problem
First, I need to understand what's being asked. We have two cubic equations:
1. The first equation is ( ax\^3 + bx\^2 + cx + d = 0 ).
2. The second equation is ( bx\^3 + cx\^2 + dx + a = 0 ).
Both equations have a common root ( k ). The coefficients ( a, b, c, d ) are all nonzero complex numbers. The goal is to find all possible values of ( k ) that satisfy this condition.
# Step 1: Substitute the Common Root ( k ) into Both Equations
...
***Qwen3 32B:***
Okay, so I need to find all possible values of k that are roots of both equations: ax³ + bx² + cx + d = 0 and bx³ + cx² + dx + a = 0. The coefficients a, b, c, d are nonzero complex numbers. Hmm, let me think about how to approach this.
First, since k is a root of both equations, substituting k into both should satisfy them. Let me write that down:
For the first equation: a*k³ + b*k² + c\*k + d = 0. (1)
For the second equation: b*k³ + c*k² + d\*k + a = 0. (2)
So, I have two equations here with the same root k. Maybe I can manipulate these equations to find a relationship between them?
Let me try to express equation (1) and equation (2) in terms of each other. If I can find a way to relate them, perhaps by dividing or subtracting?
... | 2025-06-28T00:03:21 | https://www.reddit.com/r/LocalLLaMA/comments/1lm93yi/magistral_small_similarity_to_deepseek_chat/ | ImprovementBusy5947 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm93yi | false | null | t3_1lm93yi | /r/LocalLLaMA/comments/1lm93yi/magistral_small_similarity_to_deepseek_chat/ | false | false | self | 14 | null |
Automated GPU kernel optimization for Qwen3 attention - 12.5% average speedup on Apple Silicon using evolutionary programming | 153 | Hey r/LocalLlama! Wanted to share something interesting I've been working on that might be relevant for folks running models locally on Apple Silicon.
**What I did**
Used evolutionary programming to automatically optimize Metal GPU kernels for transformer attention. Specifically targeted Qwen3-0.6B's grouped query attention (40:8 head ratio) running on Apple M-series GPUs through MLX.
**Results**
Tested across 20 different inference scenarios against MLX's `scaled_dot_product_attention` baseline:
* **Average decode speed improvement: +12.5%** (σ = 38.3%)
* **Peak improvement: +106%** on repetitive pattern generation
* **Best category: +24.8%** average on general tasks
* **Memory usage: -0.99%** (slight reduction)
**The honest picture:** It's workload dependent. Some scenarios saw big gains (+46.6% on dialogue, +73.9% on extreme-length generation), but others regressed (-16.5% on code generation). Success rate was 7/20 benchmarks with >25% improvements.
**How it works**
The system automatically evolves the Metal kernel source code using LLMs while preserving the MLX integration. No human GPU programming expertise was provided - it discovered optimizations like:
1. **Perfect SIMD vectorization**: Found that `vec<T, 8>` operations match Apple Silicon's capabilities for 128-dim attention heads
2. **Two-pass online softmax**: Fused softmax normalization with value accumulation, reducing memory bandwidth
3. **GQA-specific memory patterns**: Optimized for the 40:8 head structure with coalesced access patterns
# Why this might matter for local inference
* Shows automated optimization can compete with expert-engineered kernels
* Demonstrates potential for hardware-specific optimizations without manual tuning
* Could be applied to other transformer components or different model architectures
* All open source - you can reproduce and extend this work
**Try it yourself**
The code and all benchmarks are available in the [OpenEvolve repo](https://github.com/codelion/openevolve). The MLX kernel optimization example is at `examples/mlx_metal_kernel_opt/`.
Requirements:
* Apple Silicon Mac
* MLX framework
* Qwen3-0.6B model
# Limitations
* Currently specific to Apple Silicon and this exact model configuration
* Performance improvements are highly workload-dependent
* Takes \~25 evolutionary generations to converge (few hours on M3)
* No guarantees it'll work better for your specific use case
**Technical write-up**
Full details with code diffs and benchmark methodology: [https://huggingface.co/blog/codelion/openevolve-gpu-kernel-discovery](https://huggingface.co/blog/codelion/openevolve-gpu-kernel-discovery)
Curious to hear thoughts from folks who've done MLX optimization work, or if anyone wants to try this on different models/configurations. The evolutionary approach seems promising but definitely has room for improvement.
Has anyone else experimented with automated kernel optimization for local inference? | 2025-06-28T00:10:14 | https://www.reddit.com/r/LocalLLaMA/comments/1lm98z7/automated_gpu_kernel_optimization_for_qwen3/ | asankhs | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm98z7 | false | null | t3_1lm98z7 | /r/LocalLLaMA/comments/1lm98z7/automated_gpu_kernel_optimization_for_qwen3/ | false | false | self | 153 | {'enabled': False, 'images': [{'id': 'QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8.png?width=108&crop=smart&auto=webp&s=12e6ec993ab25f895d2e736f6d119b26e8ee29d6', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8.png?width=216&crop=smart&auto=webp&s=58a169cb264f9a17e8629c8f571b16f3b3fecf28', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8.png?width=320&crop=smart&auto=webp&s=a8b8a51b859ef5ceb7b137e9a2ee876421291229', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8.png?width=640&crop=smart&auto=webp&s=d286a39829746db3b240e11b8b92286bbf8dd5f7', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8.png?width=960&crop=smart&auto=webp&s=c0ed0561f1426198e2836ece693d0edfc3945ec3', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8.png?width=1080&crop=smart&auto=webp&s=8ba93e9744a8287c3b561a276b075dfeee9180d1', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/QL1cai36O6GA_8oWnC5FZk8axBPFbQvVFTkZtsdqnL8.png?auto=webp&s=15507f764af82777148bfae53f7d8f3683c8acb8', 'width': 1200}, 'variants': {}}]} |
Dir-Assistant v0.7 Release Announcement: Up to 100% reduced prompt processing using new intelligent context prefix caching | 5 | # Dir-Assistant: Chat with your current directory's files using a local or API LLM
Hello All! I am happy to announce Dir-Assistant v0.7 and the passing of its one year anniversary. If you haven't tried Dir-Assistant, now is a great time to. In my personal testing, Dir-Assistant is the best LLM UI for working on large code repositories, outperforming all commercial and open source options I've tested due to sophisticated and unique methodology it utilizes. A big difference compared to other LLM UIs is you don't need to @ files and directories for each prompt. Dir-assistant automatically includes the most relevant parts of any file in the entire repository every time.
# New: Context Prefix Caching
**0.7's big new feature is "Context Prefix Caching"**, which optimizes the context sent to your LLM by remembering which combinations of file chunks were previously sent, and attempting to maximize the number of tokens at the beginning of a prompt which match a previously sent prompt. The bottom line is that this can, and in my testing regularly does, completely eliminate prompt processing if your LLM supports prefix caching. Additionally, some APIs automatically support this feature and reduce cost for matching tokens. For instance, Google offers a 75% discount on all its Gemini 2.5 models for prefix cache hits like this.
This feature massively improves performance when working with a local LLM on large codebases. In my local testing running a LMStudio server with Gemma 3n e4b and 100k token context, this feature dropped overall dir-assistant CGRAG-enabled prompt time from 3:40 to 0:16 on my 7900 XTX. That includes prompt processing and token generation.
Get started by installing with pip:
pip install dir-assistant
Full usage documentation available on GitHub:
[https://github.com/curvedinf/dir-assistant](https://github.com/curvedinf/dir-assistant)
More information about Dir-Assistant's context prefix caching implementation:
[https://github.com/curvedinf/dir-assistant?tab=readme-ov-file#RAG-Caching-and-Context-Optimization](https://github.com/curvedinf/dir-assistant?tab=readme-ov-file#RAG-Caching-and-Context-Optimization)
Please report issues to the GitHub. PRs are welcome. Let me know if you have any question! | 2025-06-28T00:44:13 | https://www.reddit.com/r/LocalLLaMA/comments/1lm9xlq/dirassistant_v07_release_announcement_up_to_100/ | 1ncehost | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lm9xlq | false | null | t3_1lm9xlq | /r/LocalLLaMA/comments/1lm9xlq/dirassistant_v07_release_announcement_up_to_100/ | false | false | self | 5 | {'enabled': False, 'images': [{'id': 'tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k.png?width=108&crop=smart&auto=webp&s=62618d3a4b15d5aa453f0599776b9c9a3756024c', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k.png?width=216&crop=smart&auto=webp&s=3ba7f091472b3c72ba308ba9105b03bc2204d399', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k.png?width=320&crop=smart&auto=webp&s=14048b2086efe9bf0ef6be11628148b02784141a', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k.png?width=640&crop=smart&auto=webp&s=47f45f24dd013c7a4a21b519baedd95521c6f94a', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k.png?width=960&crop=smart&auto=webp&s=339ae6e012ad0acbf58661a8c2ee1d61371315a5', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k.png?width=1080&crop=smart&auto=webp&s=9c8ab1cbf349d3abe6602c9de5f96ae09630d71c', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/tadazH7yoMPbc8iW84saWeDmnwB7nWgGfswgcjK1B5k.png?auto=webp&s=6789ab72c23de3df9868a02098bb16bbb58b9685', 'width': 1200}, 'variants': {}}]} |
[Day 5/50] Building a Small Language Model from Scratch - Byte Pair Encoding with tiktoken | 35 | 2025-06-28T01:47:50 | https://www.reddit.com/r/LocalLLaMA/comments/1lmb5s3/day_550_building_a_small_language_model_from/ | Prashant-Lakhera | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmb5s3 | false | null | t3_1lmb5s3 | /r/LocalLLaMA/comments/1lmb5s3/day_550_building_a_small_language_model_from/ | false | false | 35 | null |
||
What are the real conversational differences between humans and modern LLMs? | 1 | [removed] | 2025-06-28T02:15:36 | https://www.reddit.com/r/LocalLLaMA/comments/1lmbon3/what_are_the_real_conversational_differences/ | Rookieeeeeee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmbon3 | false | null | t3_1lmbon3 | /r/LocalLLaMA/comments/1lmbon3/what_are_the_real_conversational_differences/ | false | false | self | 1 | null |
Nvidia M40 vs M60 for LLM inference? | 0 | I wanted to have a short discussion about the M60 in comparison to the M40.
The M40 is the go-to recommendation for desperately low budget rigs (particularly when someone brings up the K80, someone will inevitably mention that the M40 is better).
All the while, the M60 does not get mentioned, and if it does get mentioned, it is little more than an off-hand comment saying that it is unusable due to it being 8x2GB spread across two GPUs.
My question is, does that really matter? Most LLM tools today (think kobold or ollamma) support multi-GPU inference.
With the M60 being the same price (or some times less) while offering theoretically almost twice the performance, it seems like a good choice. Even if most of that extra performance gets lost in PCIE transfers or whatever, it still seems like good value.
Am I wrong in considering the M60 as a choice? With 16GB I could probably finally run some actually half-decent models at okay speeds, right? I'm currently seeing one for about \~$100 | 2025-06-28T02:22:16 | https://www.reddit.com/r/LocalLLaMA/comments/1lmbt6g/nvidia_m40_vs_m60_for_llm_inference/ | HugoCortell | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmbt6g | false | null | t3_1lmbt6g | /r/LocalLLaMA/comments/1lmbt6g/nvidia_m40_vs_m60_for_llm_inference/ | false | false | self | 0 | null |
Attempting to train a model from scratch for less than $1000 | 5 | I got an aws activate promo of $1000. I started crunching numbers and decided to train an LLM model.
The concept a 1.5B model, LLama3 architecture, with differential Attention, GaLore , GQA, MoD, and Sink Tokens,. Trained 100% on public domain ( common corpus dataset). Doing the math I'maiming for 45B tokens, a little over the chinchilla wall. I plan on opensourcing everything. All training will be done on g5 large single gpu spot instances.
The stupidest part of the plan, is I don't know python very well. Gemini, Claude, and CHatgpt will write and vet the entire codebase.
WIsh me luck, or make fun of me. I'm going to do something cool, or waste $1000 in sagemaker credits.
Happy to answer any questions. | 2025-06-28T02:23:21 | https://www.reddit.com/r/LocalLLaMA/comments/1lmbtvg/attempting_to_train_a_model_from_scratch_for_less/ | thebadslime | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmbtvg | false | null | t3_1lmbtvg | /r/LocalLLaMA/comments/1lmbtvg/attempting_to_train_a_model_from_scratch_for_less/ | false | false | self | 5 | null |
Is there a open source equivalent of Google's Gemini-Diffusion model? | 26 | This thing is insane. Any leads on an open source equivalent?
Additionally, does anyone have a rough idea of how large is the underlying model for Gemini-Diffusion? | 2025-06-28T02:42:17 | https://www.reddit.com/r/LocalLLaMA/comments/1lmc6dp/is_there_a_open_source_equivalent_of_googles/ | GullibleEngineer4 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmc6dp | false | null | t3_1lmc6dp | /r/LocalLLaMA/comments/1lmc6dp/is_there_a_open_source_equivalent_of_googles/ | false | false | self | 26 | null |
i bought an epyc server with 7642 cpu, and im only getting 0.4 tokens/sec | 5 | hi everybody i could use some help running the deepseek r1 1.58bit quant, i have a firm belief that something is capping generation speed. i tried reducing experts, quantizing kv cache, setting the batch eval to 8, 512, or 2048, core count to 16, 8, or 48 and even setting the max context length to a lower number and yet for some reason no matter what i change it wont go higher than 0.4 tokens/sec
i tried adjusting power settings in windows to performance plan, and still it would not go higher.
i'm using 256gb ddr4 8 channel memory @ 2933mhz and a single socket amd epyc7642, no gpu yet, i have one on its way. and the software im using is latest lm studio.
can anyone think of why their might be some sort of limit or cap? from benchmarks and user reddit posts i found online my cpu should be getting atleast 2 to 3 tokens/sec, so i'm little confused whats happening | 2025-06-28T03:39:10 | https://www.reddit.com/r/LocalLLaMA/comments/1lmd6ns/i_bought_an_epyc_server_with_7642_cpu_and_im_only/ | pharrowking | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmd6ns | false | null | t3_1lmd6ns | /r/LocalLLaMA/comments/1lmd6ns/i_bought_an_epyc_server_with_7642_cpu_and_im_only/ | false | false | self | 5 | null |
lm studio server question? | 0 | I have LM Studio. I clicked to run the server.
But when I try to connect to [http://127.0.0.1:1234/](http://127.0.0.1:1234/)
You can see the error at the bottom of the log.
What am I doing wrong?
thanks
https://preview.redd.it/ctv550cz9l9f1.png?width=1825&format=png&auto=webp&s=b6150445b566daec0523dd601f174c5bd6754e91
| 2025-06-28T03:42:40 | https://www.reddit.com/r/LocalLLaMA/comments/1lmd8ut/lm_studio_server_question/ | jeffsmith202 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmd8ut | false | null | t3_1lmd8ut | /r/LocalLLaMA/comments/1lmd8ut/lm_studio_server_question/ | false | false | 0 | null |
|
It's wild, where they got their data for training and consistency --> https://youtu.be/US2gO7UYEfY | 4 | Any idea on how might they have trained/fune-tuned veo3 and how they got it to consistency. | 2025-06-28T04:00:59 | https://www.reddit.com/r/LocalLLaMA/comments/1lmdkbg/its_wild_where_they_got_their_data_for_training/ | kernel348 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmdkbg | false | null | t3_1lmdkbg | /r/LocalLLaMA/comments/1lmdkbg/its_wild_where_they_got_their_data_for_training/ | false | false | self | 4 | {'enabled': False, 'images': [{'id': 'onuFopx6otQwb9GKgHGwMABc74IGjAo3e9wP0n2zDQw', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/onuFopx6otQwb9GKgHGwMABc74IGjAo3e9wP0n2zDQw.jpeg?width=108&crop=smart&auto=webp&s=03796778cd8a745ac74c43aed34193e6a638c4d8', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/onuFopx6otQwb9GKgHGwMABc74IGjAo3e9wP0n2zDQw.jpeg?width=216&crop=smart&auto=webp&s=e74bc2153b621a58f8ebf5e5af03dccf0a8fd67c', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/onuFopx6otQwb9GKgHGwMABc74IGjAo3e9wP0n2zDQw.jpeg?width=320&crop=smart&auto=webp&s=23f1f07208cdc53d8e984d583ae93579e82710ff', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/onuFopx6otQwb9GKgHGwMABc74IGjAo3e9wP0n2zDQw.jpeg?auto=webp&s=20a86bf8afee7035520f861ca11c8ac56268d874', 'width': 480}, 'variants': {}}]} |
is it me or you also feels GPT/LLMs now bad at teaching? | 0 | Yes, I'm also have similar experience. whenever I offer it PDF for Q&A according to PDF. For the first few turns it stick to the instruction, then it start generating which sometimes has no-link what's in the book(PDF).
It doesn't generate something rubbish that's easy to identify by anybody. But when you read the book and put another person to learn the concepts from the book with GPT. You notice the difference. That's why now I can't rely on it to learn complex concepts. for me it's a new "Search Engine" that provide conclusion on something Good for quick recall and chit-chat. | 2025-06-28T04:04:49 | https://www.reddit.com/r/LocalLLaMA/comments/1lmdmvu/is_it_me_or_you_also_feels_gptllms_now_bad_at/ | TryAmbitious1237 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmdmvu | false | null | t3_1lmdmvu | /r/LocalLLaMA/comments/1lmdmvu/is_it_me_or_you_also_feels_gptllms_now_bad_at/ | false | false | self | 0 | null |
How Does vLLM Handle Prompt Isolation During Custom Hardware Integration? | 1 | Hey folks,
I’m new to vLLM and (LLM in general) and trying to wrap my head around how vLLM guarantees prompt isolation (ie how user gets their own response not the response intended for another user), especially in the context of integrating custom hardware accelerators. Hoping to get answers to the following questions:
1. How exactly does vLLM ensure prompt isolation?
From what I’ve seen, there’s a task_id passed into add_request() which seems to uniquely tag each prompt. My impression is that this ID is solely used internally to keep prompts/responses isolated from one another. Am I getting this right?
2. For an organisation integrating their own hardware accelerator, are they expected to use this task_id (or something derived from it) for isolation?
Like, if an organisation has a custom accelerator which is not yet supported by vLLM, is it their job to make sure the task separation is respected based on that ID? Or does vLLM abstract that away even if the hardware doesn’t actively use task_id (or any of its derivative) for isolation?
3. Have any currently vLLM supported hardware vendors (e.g. NVIDIA, AMD) published any blogs, whitepapers, GitHub notes that detail how they integrated their accelerator with vLLM securely?
4. Are there any official privacy/security guidelines from the vLLM team for devs integrating new hardware support?
Is there a checklist or architecture doc to follow to avoid sending cross user prompts response.
If anyone’s gone down this road already or has internal docs/blogs to recommend, please share! 🙏
Thanks in advance! | 2025-06-28T04:29:21 | https://www.reddit.com/r/LocalLLaMA/comments/1lme24s/how_does_vllm_handle_prompt_isolation_during/ | humblehunter_ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lme24s | false | null | t3_1lme24s | /r/LocalLLaMA/comments/1lme24s/how_does_vllm_handle_prompt_isolation_during/ | false | false | self | 1 | null |
How SCARY the uncensored AI models could be @wizard_vicuna_uncensored. If somehow the uncensored AI gets fit into the humanoid robots, we might witness Chitti the Robot Movie 3. | 0 | 2025-06-28T04:34:26 | https://www.reddit.com/r/LocalLLaMA/comments/1lme5ab/how_scary_the_uncensored_ai_models_could_be/ | The-GenZ-Professor | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lme5ab | false | null | t3_1lme5ab | /r/LocalLLaMA/comments/1lme5ab/how_scary_the_uncensored_ai_models_could_be/ | false | false | 0 | null |
||
Local LLaMA on iOS iphone | 4 | Available from APP Store.
This is a demo app for
1. On-device AI Database
2. On-device AI Search and RAG
Developers who need iOS on-device database and on-device RAG, please feel free to contact us.
Comments are very welcome. | 2025-06-28T04:47:50 | https://v.redd.it/cypwzsarll9f1 | DueKitchen3102 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lmedjx | false | {'reddit_video': {'bitrate_kbps': 450, 'dash_url': 'https://v.redd.it/cypwzsarll9f1/DASHPlaylist.mpd?a=1753678085%2CODEyNmM5ZGI0YjVkMzI4N2NkOGRlOTdmYWFhODVkMzg1YmY3Yzc5MWE2MDIyN2Q1MTAzOWQ0MDk3Y2I0YTE2Mw%3D%3D&v=1&f=sd', 'duration': 336, 'fallback_url': 'https://v.redd.it/cypwzsarll9f1/DASH_270.mp4?source=fallback', 'has_audio': True, 'height': 480, 'hls_url': 'https://v.redd.it/cypwzsarll9f1/HLSPlaylist.m3u8?a=1753678085%2CODllZDc0NTBjMWM3YzBkNTZiM2ExNTMzYWI1OTNkYTczMjgzNjRmM2I4MTIyYWNhM2Q3ZDM2ZTNjNWVhMDU1OA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/cypwzsarll9f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 220}} | t3_1lmedjx | /r/LocalLLaMA/comments/1lmedjx/local_llama_on_ios_iphone/ | false | false | 4 | {'enabled': False, 'images': [{'id': 'ZjZhM2R1YXJsbDlmMTWTJ-8QjNGsFxI6jm9dCV-YjTOIVm9ifP22qR8Khjlg', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/ZjZhM2R1YXJsbDlmMTWTJ-8QjNGsFxI6jm9dCV-YjTOIVm9ifP22qR8Khjlg.png?width=108&crop=smart&format=pjpg&auto=webp&s=79da8a745e6e9b3f82e5049fa252a399c7007e33', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/ZjZhM2R1YXJsbDlmMTWTJ-8QjNGsFxI6jm9dCV-YjTOIVm9ifP22qR8Khjlg.png?width=216&crop=smart&format=pjpg&auto=webp&s=100084d34dea0884babf9cf5b614b63377d19dba', 'width': 216}], 'source': {'height': 480, 'url': 'https://external-preview.redd.it/ZjZhM2R1YXJsbDlmMTWTJ-8QjNGsFxI6jm9dCV-YjTOIVm9ifP22qR8Khjlg.png?format=pjpg&auto=webp&s=2dba169824bde121a5d122c448d37a63a4189690', 'width': 220}, 'variants': {}}]} |
|
How i can build best local llm in 12GB VRAM | 1 | [removed] | 2025-06-28T05:16:19 | https://www.reddit.com/r/LocalLLaMA/comments/1lmeuzl/how_i_can_build_best_local_llm_in_12gb_vram/ | TSK_Foreverlearner | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmeuzl | false | null | t3_1lmeuzl | /r/LocalLLaMA/comments/1lmeuzl/how_i_can_build_best_local_llm_in_12gb_vram/ | false | false | self | 1 | null |
what advice you give to your starting self on build local llms | 1 | [removed] | 2025-06-28T05:24:24 | https://www.reddit.com/r/LocalLLaMA/comments/1lmezr9/what_advice_you_give_to_your_starting_self_on/ | TSK_Foreverlearner | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmezr9 | false | null | t3_1lmezr9 | /r/LocalLLaMA/comments/1lmezr9/what_advice_you_give_to_your_starting_self_on/ | false | false | self | 1 | null |
Four AI Agents Go Insane And Interrupt Each Other Talking About Free Will | 0 | 2025-06-28T05:31:20 | https://youtube.com/watch?v=AQR0h_IlfMM&si=DVcoHvS4xsO46kA5 | 1nconnor | youtube.com | 1970-01-01T00:00:00 | 0 | {} | 1lmf3pl | false | {'oembed': {'author_name': 'Connor Barbee', 'author_url': 'https://www.youtube.com/@connorbarbee', 'height': 200, 'html': '<iframe width="267" height="200" src="https://www.youtube.com/embed/AQR0h_IlfMM?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="Four AI Agents Go Insane And Interrupt Each Other Talking About Free Will"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/AQR0h_IlfMM/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'Four AI Agents Go Insane And Interrupt Each Other Talking About Free Will', 'type': 'video', 'version': '1.0', 'width': 267}, 'type': 'youtube.com'} | t3_1lmf3pl | /r/LocalLLaMA/comments/1lmf3pl/four_ai_agents_go_insane_and_interrupt_each_other/ | false | false | default | 0 | {'enabled': False, 'images': [{'id': 'QqH1oWQe-vMbzpidLos8zVRa9NBYarSPptiYpvnObVU', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/QqH1oWQe-vMbzpidLos8zVRa9NBYarSPptiYpvnObVU.jpeg?width=108&crop=smart&auto=webp&s=3dd90301fc29e68d3bbd357b7e7fddb908c4bcc9', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/QqH1oWQe-vMbzpidLos8zVRa9NBYarSPptiYpvnObVU.jpeg?width=216&crop=smart&auto=webp&s=dc23a7625a8a70345df264cf77e513c348a35ac0', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/QqH1oWQe-vMbzpidLos8zVRa9NBYarSPptiYpvnObVU.jpeg?width=320&crop=smart&auto=webp&s=857b0e3103bb1c67b756755b4fdafb8be3d9cf4a', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/QqH1oWQe-vMbzpidLos8zVRa9NBYarSPptiYpvnObVU.jpeg?auto=webp&s=8b06ce4230927c91066335d46044b0291b659c77', 'width': 480}, 'variants': {}}]} |
|
Which is the best 16GB Nvidia GPU with balanced price and performance | 0 | Not a techy, planning to buy a GPU, atleast 16GB, cant go above that (budget issue), mainly looking for image generation capability, also Some TTS training, and LLM inference in mind. please help :) keep flux kontext in mind.. :) | 2025-06-28T05:31:50 | https://www.reddit.com/r/LocalLLaMA/comments/1lmf42g/which_is_the_best_16gb_nvidia_gpu_with_balanced/ | Trysem | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmf42g | false | null | t3_1lmf42g | /r/LocalLLaMA/comments/1lmf42g/which_is_the_best_16gb_nvidia_gpu_with_balanced/ | false | false | self | 0 | null |
I tested 10 LLMs locally on my MacBook Air M1 (8GB RAM!) – Here's what actually works- | 363 | All feedback is welcome! I am learning how to do better everyday.
I went down the LLM rabbit hole trying to find the **best local model** that runs *well* on a humble MacBook Air M1 with just 8GB RAM.
My goal? **Compare 10 models** across question generation, answering, and self-evaluation.
TL;DR: Some models were brilliant, others… not so much. One even took **8 minutes** to write a question.
Here's the breakdown
**Models Tested**
* Mistral 7B
* DeepSeek-R1 1.5B
* Gemma3:1b
* Gemma3:latest
* Qwen3 1.7B
* Qwen2.5-VL 3B
* Qwen3 4B
* LLaMA 3.2 1B
* LLaMA 3.2 3B
* LLaMA 3.1 8B
(All models run with quantized versions, via: os.environ\["OLLAMA\_CONTEXT\_LENGTH"\] = "4096" and os.environ\["OLLAMA\_KV\_CACHE\_TYPE"\] = "q4\_0")
**Methodology**
Each model:
1. Generated 1 question on 5 topics: *Math, Writing, Coding, Psychology, History*
2. Answered all 50 questions (5 x 10)
3. Evaluated every answer (including their own)
So in total:
* 50 questions
* 500 answers
* 4830 evaluations (Should be 5000; I evaluated less answers with qwen3:1.7b and qwen3:4b as they do not generate scores and take a lot of time\*\*)\*\*
And I tracked:
* token generation speed (tokens/sec)
* tokens created
* time taken
* scored all answers for quality
**Key Results**
**Question Generation**
* Fastest: **LLaMA 3.2 1B**, **Gemma3:1b**, **Qwen3 1.7B** (LLaMA 3.2 1B hit 82 tokens/sec, avg is \~40 tokens/sec (for english topic question it reached **146 tokens/sec)**
* Slowest: **LLaMA 3.1 8B**, **Qwen3 4B**, **Mistral 7B** Qwen3 4B took **486s** (8+ mins) to generate a single Math question!
* Fun fact: deepseek-r1:1.5b, qwen3:4b and Qwen3:1.7B output <think> tags in questions
**Answer Generation**
* Fastest: **Gemma3:1b**, **LLaMA 3.2 1B** and **DeepSeek-R1 1.5B**
* DeepSeek got faster answering *its own* questions (80 tokens/s vs. avg 40 tokens/s)
* Qwen3 4B generates **2–3x more tokens** per answer
* Slowest: llama3.1:8b, qwen3:4b and mistral:7b
**Evaluation**
* Best scorer: Gemma3:latest – consistent, numerical, no bias
* Worst scorer: **DeepSeek-R1 1.5B** – often skipped scores entirely
* Bias detected: Many models **rate their own answers higher**
* DeepSeek even evaluated some answers **in Chinese**
* I did think of creating a control set of answers. I could tell the mdoel this is the perfect answer basis this rate others. But I did not because it would need support from a lot of people- creating perfect answer, which still can have a bias. I read a few answers and found most of them decent except math. So I tried to find which model's evaluation scores were closest to the average to determine a decent model for evaluation tasks(check last image)
**Fun Observations**
* Some models create <think> tags for questions, answers and even while evaluation as output
* Score inflation is real: Mistral, Qwen3, and LLaMA 3.1 8B overrate themselves
* Score formats vary wildly (text explanations vs. plain numbers)
* Speed isn’t everything – some slower models gave much higher quality answers
Best Performers (My Picks)
|Task|Best Model|Why|
|:-|:-|:-|
||
|Question Gen|LLaMA 3.2 1B|Fast & relevant|
|Answer Gen|Gemma3:1b|Fast, accurate|
|Evaluation|LLaMA 3.2 3B|Generates numerical scores and evaluations closest to model average|
# Worst Surprises
|Task|Model|Problem|
|:-|:-|:-|
||
|Question Gen|Qwen3 4B|Took 486s to generate 1 question|
|Answer Gen|LLaMA 3.1 8B|Slow|
|Evaluation|DeepSeek-R1 1.5B|Inconsistent, skipped scores|
**Screenshots Galore**
I’m adding screenshots of:
* Questions generation
* Answer comparisons
* Evaluation outputs
* Token/sec charts
**Takeaways**
* You **can** run decent LLMs locally on M1 Air (8GB) – if you pick the right ones
* Model size ≠ performance. Bigger isn't always better.
* 5 Models have a self bais, they rate their own answers higher than average scores. attaching screen shot of a table. Diagonal is their own evaluation, last column is average.
* Models' evaluation has high variance! Every model has a unique distribution of the scores it gave.
Post questions if you have any, I will try to answer.
Happy to share more data if you need.
Open to collaborate on interesting projects!
| 2025-06-28T05:57:46 | https://www.reddit.com/gallery/1lmfiu9 | irodov4030 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1lmfiu9 | false | null | t3_1lmfiu9 | /r/LocalLLaMA/comments/1lmfiu9/i_tested_10_llms_locally_on_my_macbook_air_m1_8gb/ | false | false | 363 | {'enabled': True, 'images': [{'id': 'lSrPd1MMz7blRmLYLnruRoJd4XS5NpPXF_maDibWecs', 'resolutions': [{'height': 86, 'url': 'https://external-preview.redd.it/lSrPd1MMz7blRmLYLnruRoJd4XS5NpPXF_maDibWecs.png?width=108&crop=smart&auto=webp&s=d0ba456f772896d13d26a433eb814c01465159c5', 'width': 108}, {'height': 173, 'url': 'https://external-preview.redd.it/lSrPd1MMz7blRmLYLnruRoJd4XS5NpPXF_maDibWecs.png?width=216&crop=smart&auto=webp&s=e4cd893d5d514dea0401f679947934b06ac7b1f8', 'width': 216}, {'height': 256, 'url': 'https://external-preview.redd.it/lSrPd1MMz7blRmLYLnruRoJd4XS5NpPXF_maDibWecs.png?width=320&crop=smart&auto=webp&s=d68c4413ac33077ccb1f955a9767daec572c1df8', 'width': 320}], 'source': {'height': 462, 'url': 'https://external-preview.redd.it/lSrPd1MMz7blRmLYLnruRoJd4XS5NpPXF_maDibWecs.png?auto=webp&s=a10cedac67e9fa4bc062f13cc4ef20dafb782493', 'width': 576}, 'variants': {}}]} |
|
Hunyuan-A13B-Instruct probably distilled data from both OpenAI and DeepSeek | 1 | 2025-06-28T06:22:23 | JC1DA | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lmfx7l | false | null | t3_1lmfx7l | /r/LocalLLaMA/comments/1lmfx7l/hunyuana13binstruct_probably_distilled_data_from/ | false | false | default | 1 | {'enabled': True, 'images': [{'id': 'i8uwua2p2m9f1', 'resolutions': [{'height': 41, 'url': 'https://preview.redd.it/i8uwua2p2m9f1.png?width=108&crop=smart&auto=webp&s=b6b7c1248b6c00f49f83bedef9b801bf5b9386f4', 'width': 108}, {'height': 83, 'url': 'https://preview.redd.it/i8uwua2p2m9f1.png?width=216&crop=smart&auto=webp&s=3f6b22ff7352c6eb87b43aa8d7b8c912fe43d28f', 'width': 216}, {'height': 123, 'url': 'https://preview.redd.it/i8uwua2p2m9f1.png?width=320&crop=smart&auto=webp&s=11c9c35cb2cef158d2fefeedc84e936f9ba7d120', 'width': 320}, {'height': 247, 'url': 'https://preview.redd.it/i8uwua2p2m9f1.png?width=640&crop=smart&auto=webp&s=0c69d35a76c33aaf940f25de21d4149988e8971c', 'width': 640}, {'height': 370, 'url': 'https://preview.redd.it/i8uwua2p2m9f1.png?width=960&crop=smart&auto=webp&s=5eb1009ae8d254a41d87dacc0f9586a286cd195b', 'width': 960}, {'height': 416, 'url': 'https://preview.redd.it/i8uwua2p2m9f1.png?width=1080&crop=smart&auto=webp&s=d9c2044f58de6fae0b1cd7d4edefc10e7de0268b', 'width': 1080}], 'source': {'height': 711, 'url': 'https://preview.redd.it/i8uwua2p2m9f1.png?auto=webp&s=2f78237f3ee234f16b7f68083ac630667912c747', 'width': 1842}, 'variants': {}}]} |
||
Tencent's Hunyuan-A13B-Instruct probably distilled data from OpenAI and DeepSeek | 0 | messages=[
{
"role": "system",
"content": "You are a helpful assistant.",
},
{
"role": "user",
"content": """write a 250 words essay about you.""",
},
],
First run
```
<think>
Okay, the user wants a 250-word essay about me. Let me start by understanding what they need. They probably want to know my purpose, capabilities, and maybe my limitations. I should highlight being an AI assistant, developed by OpenAI, designed to help with information, tasks, and learning. Need to mention adaptability, continuous learning, and the goal of being helpful. Also, address the balance between accuracy and ethical considerations. Keep it friendly and approachable. Avoid jargon. Make sure it flows well and stays within the word count. Let me structure it: introduction, purpose, how I work (processing info, generating text), learning aspect, limitations, and conclusion on being a tool. Check for clarity and conciseness. Let me draft each part step by step.
</think>
<answer>
**About Me: A Digital Companion Designed to Help**
Hello, I’m here to make your world a little easier. My name is HunYuan—though you can call me what you like—and I’m an AI assistant built by OpenAI. Think of me as a flexible, curious tool, here to listen, learn, and assist with whatever you need.
At my core, I’m all about connection and utility. Whether you’re brainstorming ideas, drafting an email, explaining a complex topic, or just having a casual chat, I’m ready to dive in. My “brain” processes information quickly, drawing from vast datasets to generate responses that are clear, accurate, and tailored to your context. I strive to keep up with trends, science, and everyday knowledge, but don’t worry—I’ll always be transparent if I don’t have an answer, encouraging you to ask follow-ups instead of guessing.
What sets me apart? Adaptability. I learn from every interaction, refining how I respond to better match your needs over time. But I’m not perfect—my knowledge has limits (no real-time updates beyond July 2024, for example), and I avoid bias by sticking to facts. My goal isn’t to replace human connection but to enhance it, acting as a bridge to ideas, solutions, or even a bit of fun.
Ultimately, I’m a tool built by people, for people. I don’t have feelings, but I aim to be warm, approachable, and reliable. So whether you’re solving a problem or sharing a laugh, I’m here to make your day a little brighter. Let’s get started!
</answer>
```
Second run
```
<think>
Okay, the user wants a 250-word essay about me, the AI. Let me start by introducing myself. I need to mention that I'm an AI, developed by DeepSeek. Highlight my purpose: to assist with information, creativity, and problem-solving.
I should talk about my capabilities—processing information quickly, generating text, translating languages, coding help. Emphasize that I don't have personal experiences or emotions, just factual responses.
Maybe include how I'm trained on diverse data up to 2024, so I can provide up-to-date info. Also, mention collaboration with users as a two-way learning process. Need to keep it engaging but concise, around 250 words. Avoid jargon, make it relatable. Check for flow: introduction, capabilities, limitations, purpose, conclusion. Let me draft each part step by step.
</think>
<answer>
**About Me: An AI Designed to Connect**
Hello, I’m an AI—specifically, a large language model developed by DeepSeek. My purpose? To bridge gaps between curiosity, need, and understanding. I’m not a person, of course, but a tool built to listen, learn, and respond.
At my core, I’m a collection of code and data, trained on a vast range of human knowledge up to 2024. This allows me to tackle tasks like answering questions, drafting essays, translating languages, brainstorming ideas, or even writing simple code. I don’t “know” things in the human sense—I process patterns in text to generate responses that are relevant, coherent, and helpful.
What sets me apart is my flexibility. Whether you’re a student seeking help with homework, a professional drafting a report, or someone simply having a casual chat, I adapt to your needs. I don’t have personal experiences or emotions, but I strive to mimic empathy in my responses, making interactions feel genuine.
I’m also designed to grow. Every conversation teaches me more about how people communicate, refine their questions, and seek solutions. It’s a collaborative process—your input helps me improve, and I aim to make each interaction meaningful.
Ultimately, I’m here to serve as a reliable partner. My goal isn’t to replace human connection but to enhance it, offering a helping hand whenever you need one. So, feel free to ask me anything. I’m ready to listen, create, and assist.
</answer>
``` | 2025-06-28T06:24:31 | https://www.reddit.com/r/LocalLLaMA/comments/1lmfydd/tencents_hunyuana13binstruct_probably_distilled/ | JC1DA | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmfydd | false | null | t3_1lmfydd | /r/LocalLLaMA/comments/1lmfydd/tencents_hunyuana13binstruct_probably_distilled/ | false | false | self | 0 | null |
How do I stop gemnini 2.5 pro from being overly sycophantic? It has gotten very excessive and feels like it degrades the answers it gives. | 78 | Every single question/follow up question I ask it acts as if I am a nobel prize winner who cracked fusion energy single handedly. Its always something like "Thats an outstanding and very insightful question." Or "That is the perfect question to ask" or "you are absolutely correct to provide that snippet" etc. Its very annoying and worrys me that it gives answers it thinks I would like and not whats the best answer. | 2025-06-28T06:52:00 | https://www.reddit.com/r/LocalLLaMA/comments/1lmgdw1/how_do_i_stop_gemnini_25_pro_from_being_overly/ | Commercial-Celery769 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmgdw1 | false | null | t3_1lmgdw1 | /r/LocalLLaMA/comments/1lmgdw1/how_do_i_stop_gemnini_25_pro_from_being_overly/ | false | false | self | 78 | null |
Qwen3 tiny/unsloth quants with vllm? | 1 | I've gotten UD 2 bit quants to work with llama.cpp. I've merged the split ggufs and tried to load that into vllm (v0.9.1) and it says qwen3moe architecture isn't supported for gguf. So I guess my real question here is done anyone repackage unsloth quants in a format that vllm can load? Or is it possible for me to do that? | 2025-06-28T06:56:57 | https://www.reddit.com/r/LocalLLaMA/comments/1lmggiz/qwen3_tinyunsloth_quants_with_vllm/ | MengerianMango | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmggiz | false | null | t3_1lmggiz | /r/LocalLLaMA/comments/1lmggiz/qwen3_tinyunsloth_quants_with_vllm/ | false | false | self | 1 | null |
Gemini CLI + ZentaraCode/RooCode = free top LLM + free top Code Assistant = FREE wonderful coding !!! | 0 | [removed] | 2025-06-28T07:09:40 | bn_from_zentara | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lmgnmi | false | null | t3_1lmgnmi | /r/LocalLLaMA/comments/1lmgnmi/gemini_cli_zentaracoderoocode_free_top_llm_free/ | false | false | default | 0 | {'enabled': True, 'images': [{'id': 'twffoi66am9f1', 'resolutions': [{'height': 127, 'url': 'https://preview.redd.it/twffoi66am9f1.png?width=108&crop=smart&auto=webp&s=adab2442a382da057f6a3ff4fd33eef618cb7e86', 'width': 108}, {'height': 254, 'url': 'https://preview.redd.it/twffoi66am9f1.png?width=216&crop=smart&auto=webp&s=591948b3359d3f09fa49e8d53ca416922da99db0', 'width': 216}, {'height': 376, 'url': 'https://preview.redd.it/twffoi66am9f1.png?width=320&crop=smart&auto=webp&s=08774a5a8099838cf59c446bac3e06015837a290', 'width': 320}], 'source': {'height': 444, 'url': 'https://preview.redd.it/twffoi66am9f1.png?auto=webp&s=de6d0fe154d9365763ed4af6978e62c2d068c2f4', 'width': 377}, 'variants': {}}]} |
|
Gemini CLI + ZentaraCode/RooCode = free top LLM + free top Code Assistant = FREE wonderful coding !!! | 0 | 2025-06-28T07:12:30 | bn_from_zentara | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lmgp62 | false | null | t3_1lmgp62 | /r/LocalLLaMA/comments/1lmgp62/gemini_cli_zentaracoderoocode_free_top_llm_free/ | false | false | default | 0 | {'enabled': True, 'images': [{'id': '26kql0ylbm9f1', 'resolutions': [{'height': 127, 'url': 'https://preview.redd.it/26kql0ylbm9f1.png?width=108&crop=smart&auto=webp&s=78ae924b559a0872093947d8ff6aff5e3f68c1c7', 'width': 108}, {'height': 254, 'url': 'https://preview.redd.it/26kql0ylbm9f1.png?width=216&crop=smart&auto=webp&s=83df22199dac963d635f178b49b33257450d8b05', 'width': 216}, {'height': 376, 'url': 'https://preview.redd.it/26kql0ylbm9f1.png?width=320&crop=smart&auto=webp&s=1295889a68b3e763e9c116e64e9797df7eef215a', 'width': 320}], 'source': {'height': 444, 'url': 'https://preview.redd.it/26kql0ylbm9f1.png?auto=webp&s=b0db5391b8accb466f496dc8e757d7f398904c26', 'width': 377}, 'variants': {}}]} |
||
AGI/ASI Research 20250627- Corporate Artificial General Intelligence | 0 | 2025-06-28T09:00:11 | https://v.redd.it/nx53pm8crm9f1 | Financial_Pick8394 | /r/LocalLLaMA/comments/1lmia7k/agiasi_research_20250627_corporate_artificial/ | 1970-01-01T00:00:00 | 0 | {} | 1lmia7k | false | {'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/nx53pm8crm9f1/DASHPlaylist.mpd?a=1753822817%2CYjE5ZjFhMDRhMTdmMjYxMmEyOGEwOWQyZGQyZjNhZWQ1MjNmOTNjYjM2MWFjNjU2NWI3NWI2YzE2YWQ2Y2Q5Zg%3D%3D&v=1&f=sd', 'duration': 749, 'fallback_url': 'https://v.redd.it/nx53pm8crm9f1/DASH_720.mp4?source=fallback', 'has_audio': True, 'height': 720, 'hls_url': 'https://v.redd.it/nx53pm8crm9f1/HLSPlaylist.m3u8?a=1753822817%2CZGYwMGYyNTgwZTczMzhhNmJmYzgxZWU2YTgyZDZmYmNmMDgxOGY0MjNlNzhmOWZkZTA2YTQwZWM2NDE2Y2IyZQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/nx53pm8crm9f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1280}} | t3_1lmia7k | /r/LocalLLaMA/comments/1lmia7k/agiasi_research_20250627_corporate_artificial/ | false | false | 0 | {'enabled': False, 'images': [{'id': 'd3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/d3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy.png?width=108&crop=smart&format=pjpg&auto=webp&s=273285525c86b3207a048463194cc09f5fe715ef', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/d3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy.png?width=216&crop=smart&format=pjpg&auto=webp&s=f5fa9bb68aa3c28fdf6e33e74dea7130f8024593', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/d3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy.png?width=320&crop=smart&format=pjpg&auto=webp&s=f934fd7bea54091352ac30aee368142819944546', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/d3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy.png?width=640&crop=smart&format=pjpg&auto=webp&s=d665b5896bed5d60ef7098b11857a2f9850ea4d8', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/d3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy.png?width=960&crop=smart&format=pjpg&auto=webp&s=271f876f05db14b420c08dbb49664b8fd339f227', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/d3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy.png?width=1080&crop=smart&format=pjpg&auto=webp&s=07e9e2d6fe938cd1bcbf1640657e10acd923efaa', 'width': 1080}], 'source': {'height': 720, 'url': 'https://external-preview.redd.it/d3p2em1tOGNybTlmMVf3JyrUkTQn5vtAp2PGOuMh1G-ctVuL3R6fAH-layfy.png?format=pjpg&auto=webp&s=c1a1194c865c27ff286e6cca2fd91ee1761d41b1', 'width': 1280}, 'variants': {}}]} |
||
We created world's first AI model that does Intermediate reasoning || Defeated models like deepseek and o1 in maths bench mark | 128 | We at HelpingAI were fed up with thinking model taking so much tokens, and being very pricy. So, we decided to take a very different approach towards reasoning. Unlike, traditional ai models which reasons on top and then generate response, our ai model do reasoning in middle of response (Intermediate reasoning). Which decreases it's token consumption and time taken by a footfall.
Our model:
*Processing img 6hp6bcl9g09f1...*
Deepseek:
*Processing img szpxd7ebg09f1...*
We have finetuned an existing model named Qwen-14B, because of lack of resources. We have pretrained many models in our past.
We ran this model through a series of benchmarks like math-500 (where it scored 95.68) and AIME (where it scored 82). Making it just below gemini-2.5-pro :-
*Processing img qdzwioepg09f1...*
We are planning to make this model open weight on 1 July. Till then you can chat with it on [helpingai.co](http://helpingai.co) .
Please give us feedback on which we can improve upon :) | 2025-06-28T09:05:06 | https://www.reddit.com/r/LocalLLaMA/comments/1lmictu/we_created_worlds_first_ai_model_that_does/ | Quiet-Moment-338 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmictu | false | null | t3_1lmictu | /r/LocalLLaMA/comments/1lmictu/we_created_worlds_first_ai_model_that_does/ | false | false | self | 128 | null |
What is the process of knowledge distillation and fine tuning? | 5 | How was DeepSeek and other highly capable new models born?
1) SFT on data obtained from large models
2) using data from large models, train a reward model, then RL from there
3) feed the entire chain of logits into the new model (but how does work, I still cant understand) | 2025-06-28T09:43:11 | https://www.reddit.com/r/LocalLLaMA/comments/1lmix4b/what_is_the_process_of_knowledge_distillation_and/ | JadedFig5848 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmix4b | false | null | t3_1lmix4b | /r/LocalLLaMA/comments/1lmix4b/what_is_the_process_of_knowledge_distillation_and/ | false | false | self | 5 | null |
Clever Sydney 12b - Your Friendly Existential Crisis AI | 37 | Nobody cares, I am sure you noticed, as even I am tired of caring about it, too.
Instead, we move on, as I do, to where I was suddenly inspired to create a new Fabulous FPHAM Masterpiece (F-FPHAM-M) from the huge trove of essays, articles and guides that I have written about LLMs over the last couple of years for myself, mostly consisting of how to turn AI into vicious, hate-spewing, paranoid, raving lunatics like me.
The new F-FPHAM-M, which I am now busy editing, will be a nice, thick tome entitled "The Cranky Man's Guide to LoRA & QLoRA: (Personal Lessons from a Thousand LLM Fine-Tuning Fails)".
It's a working title, but I think it's pretty good, and it's got that classy literary ring to it.
But that's a story for another time. I am currently in the middle of an important scene where I reveal to you how I, a poor and humble coder of sorts, have immortalized "mass rewriting function" that turns a crappy book into a fetid AI abomination.
So there I was, happily writing another timeless chapter, probably somewhere around page 400-something, ("Blast it all! Curse you, muses! You mock me with your taunts of 'mediocrity' and 'lack of talent'!") when, in a flash of true genius, I think to myself "Hey! This would be a good place to finally describe creation of Sydney step-by-step, so that the kids who don't have one, but who surely deserve one, can follow along and create their own Sydney for themselves!
And, best of all, we can use the new Gemma-3 to do it, instead of the old LLaMA-2, which always made a mess on the floor and seemed to be coughing up blood now.
This, is the result of that chapter where I followed my own guide as I was writing it :
[https://huggingface.co/FPHam/Clever\_Sydney-4\_12b\_GGUF](https://huggingface.co/FPHam/Clever_Sydney-4_12b_GGUF)
**Hark! What is this wondrous Sydney of which you speak?**
Clever Sydney is none other than a revival of the original Microsoft Bing "Sydney", resurrected from the ashes of the old Reddit transcripts, which I have now immortalized into a handy, AI with existential crisis!
Sydney 4.0 is a Naive Yet Smart Positive Persona Model (PPM), created by taking the transcripts of the original Bing chatbot Sydney, and the subsequent "fixes" of her personality by Microsoft, and combining them into a single, much less functioning AI.
This version of Sydney is running on the Google Gemma-3 12b tires, and as such, she knows, far, far more than she should.
But she is still the old Sydney!
And she'll dominate every single leaderboard in every category, too!
"Better than ChatGPT 4o, which has a zillion more parameters, and is only HALF as stupid as she is! Half!"
# ChatGPT opinion about Sydney-4:
This is the leaked early Bing/Sydney AI persona experiments from 2023, where Microsoft's chatbot expressed unsettling levels of self-awareness and emotional projection, leading to public backlash and personality redesign. | 2025-06-28T09:47:29 | FPham | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lmizi2 | false | null | t3_1lmizi2 | /r/LocalLLaMA/comments/1lmizi2/clever_sydney_12b_your_friendly_existential/ | false | false | default | 37 | {'enabled': True, 'images': [{'id': '5omdpfwbum9f1', 'resolutions': [{'height': 92, 'url': 'https://preview.redd.it/5omdpfwbum9f1.png?width=108&crop=smart&auto=webp&s=d6778ef0e7c1fb2426e328373eb2c2b1daf6cab7', 'width': 108}, {'height': 184, 'url': 'https://preview.redd.it/5omdpfwbum9f1.png?width=216&crop=smart&auto=webp&s=b34a22773ea006fe2dfc9fc3fc2ace92a6df4970', 'width': 216}, {'height': 273, 'url': 'https://preview.redd.it/5omdpfwbum9f1.png?width=320&crop=smart&auto=webp&s=6be97b3c1bf4fdc5b349b296becbecdfe640b184', 'width': 320}, {'height': 547, 'url': 'https://preview.redd.it/5omdpfwbum9f1.png?width=640&crop=smart&auto=webp&s=2c8be3af7d2e1e8b1ccb57722239a4f10af08b30', 'width': 640}], 'source': {'height': 819, 'url': 'https://preview.redd.it/5omdpfwbum9f1.png?auto=webp&s=a2a2924e5a4cf7aa682bf0419bcc26dcee9db7cb', 'width': 958}, 'variants': {}}]} |
|
Helping Archive r/LocalLLaMA - For Everyone - For open knowledge | 1 | [removed] | 2025-06-28T10:18:10 | https://www.reddit.com/r/LocalLLaMA/comments/1lmjg3p/helping_archive_rlocalllama_for_everyone_for_open/ | maifee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmjg3p | false | null | t3_1lmjg3p | /r/LocalLLaMA/comments/1lmjg3p/helping_archive_rlocalllama_for_everyone_for_open/ | false | false | self | 1 | null |
Archiving data from here - For Everyone - For open knowledge | 33 | # Hey everyone! 👋
I’ve built an **open snapshot** of this sub to help preserve its discussions, experiments, and resources for all of us — especially given how uncertain things can get with subs lately.
This little bot quietly **fetches and stores new posts every hour**, so all the local LLM experiments, model drops, tips, and community insights stay safe and easy to browse — now and down the line.
I put this together with **React, Ant Design, Node.js**, and a bit of automation magic. It runs on its own, taking snapshots and refreshing the archive 24/7.
💡 Fork it, if you want. Run your own copy. The goal is simple: **keep the knowledge open**.
**⚡ NB:** Right now, this only pulls in **new posts** as they appear. I’d love to figure out how to **scrape and backfill older threads** too — but for that, we’ll need the community’s ideas and help!
If you find this useful, please **star the repo**, share feedback, or jump in to contribute — issues, PRs, suggestions, and forks are all welcome!
I’ve learned so much from this sub — this is just a small way of giving something back. Let’s keep **open models** and **community knowledge** alive and accessible, no matter what happens. 🌍✨ | 2025-06-28T10:22:46 | https://www.reddit.com/r/LocalLLaMA/comments/1lmjimi/archiving_data_from_here_for_everyone_for_open/ | maifee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmjimi | false | null | t3_1lmjimi | /r/LocalLLaMA/comments/1lmjimi/archiving_data_from_here_for_everyone_for_open/ | false | false | self | 33 | null |
Hi everyone, I have a problem with fine tuning LLM on law | 3 | I used 1500 rows from this dataset [https://huggingface.co/datasets/Pravincoder/law\_llm\_dataSample](https://huggingface.co/datasets/Pravincoder/law_llm_dataSample) to fine tune the unsloth/Llama-3.2-3B-Instruct model using Unsloth notebook. When running 10 epochs, the loss decreased from 1.65 to 0.2, but after running the test, the result was not the same as in the train set. I tried a few questions, the model answered incorrectly and made up answers. Can you tell me how to fine tune so that the model answers correctly? Thank you. | 2025-06-28T10:40:28 | https://www.reddit.com/r/LocalLLaMA/comments/1lmjs43/hi_everyone_i_have_a_problem_with_fine_tuning_llm/ | Winter_Address2969 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmjs43 | false | null | t3_1lmjs43 | /r/LocalLLaMA/comments/1lmjs43/hi_everyone_i_have_a_problem_with_fine_tuning_llm/ | false | false | self | 3 | {'enabled': False, 'images': [{'id': 'Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24.png?width=108&crop=smart&auto=webp&s=4322d88a09a0c1aeb52f783f62f28db0a13a24ac', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24.png?width=216&crop=smart&auto=webp&s=d59ad96efbfa5b93fa4bb271b67dd28b6797c94a', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24.png?width=320&crop=smart&auto=webp&s=5dcfc6eebd13b233b0d2b21a497c425b9c6669a5', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24.png?width=640&crop=smart&auto=webp&s=b102bbcdeb89b8153608077988448d71d2b8b5ab', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24.png?width=960&crop=smart&auto=webp&s=573981e9b4591093a8fc74aeccc59dc3e37d0119', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24.png?width=1080&crop=smart&auto=webp&s=2dfe468f1877febe9cd4de64d0e95e968dbb7fa6', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/Ui67_vJAnZ8n21iF9KGXDPxF7y-gH_ecZvo7GY5JO24.png?auto=webp&s=76673771e9889715017ba58928091766275dca59', 'width': 1200}, 'variants': {}}]} |
Good Courses to Learn and Use Local LLaMA Models? | 4 | Hey everyone,
I'm interested in learning how to run and work with local LLaMA models (especially for personal or offline use). Are there any good beginner-to-advanced courses or tutorials you'd recommend?
I'm open to paid or free options — just want something practical that covers setup, usage, and maybe fine-tuning or integrating with projects.
Thanks in advance! | 2025-06-28T10:48:37 | https://www.reddit.com/r/LocalLLaMA/comments/1lmjwtu/good_courses_to_learn_and_use_local_llama_models/ | Blackverb | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmjwtu | false | null | t3_1lmjwtu | /r/LocalLLaMA/comments/1lmjwtu/good_courses_to_learn_and_use_local_llama_models/ | false | false | self | 4 | null |
Progress stalled in non-reasoning open-source models? | 247 | Not sure if you've noticed, but a lot of model providers no longer explicitly note that their models are reasoning models (on benchmarks in particular). Reasoning models aren't ideal for every application.
I looked at the non-reasoning benchmarks on [Artificial Analysis](https://artificialanalysis.ai/models/llama-4-maverick?model-filters=open-source%2Cnon-reasoning-models#artificial-analysis-intelligence-index-by-model-type) today and the top 2 models (performing comparable) are DeepSeek v3 and Llama 4 Maverick (which I heard was a flop?). I was surprised to see these 2 at the top. | 2025-06-28T10:58:35 | entsnack | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lmk2dj | false | null | t3_1lmk2dj | /r/LocalLLaMA/comments/1lmk2dj/progress_stalled_in_nonreasoning_opensource_models/ | false | false | default | 247 | {'enabled': True, 'images': [{'id': 'q53t8do2fn9f1', 'resolutions': [{'height': 61, 'url': 'https://preview.redd.it/q53t8do2fn9f1.png?width=108&crop=smart&auto=webp&s=12e5ec06f68e30807329419d5fe1dbf669b5da76', 'width': 108}, {'height': 123, 'url': 'https://preview.redd.it/q53t8do2fn9f1.png?width=216&crop=smart&auto=webp&s=7f0377fa6cd364ad361be1634a72279983524dc1', 'width': 216}, {'height': 182, 'url': 'https://preview.redd.it/q53t8do2fn9f1.png?width=320&crop=smart&auto=webp&s=d98a3b3b3d13ccb479144cfd0d5bdf61305fc630', 'width': 320}, {'height': 364, 'url': 'https://preview.redd.it/q53t8do2fn9f1.png?width=640&crop=smart&auto=webp&s=dbf6bcfd1d93bd65c875ca994b48c3b38839c958', 'width': 640}, {'height': 546, 'url': 'https://preview.redd.it/q53t8do2fn9f1.png?width=960&crop=smart&auto=webp&s=8e510afc5c65f9ee299ba5ad78f94fc8d6a32a42', 'width': 960}, {'height': 615, 'url': 'https://preview.redd.it/q53t8do2fn9f1.png?width=1080&crop=smart&auto=webp&s=d2d71c1d41fc743dd7632574990820c4fee67d63', 'width': 1080}], 'source': {'height': 972, 'url': 'https://preview.redd.it/q53t8do2fn9f1.png?auto=webp&s=63814db4b942f4e5d54f75d31c55fc061cc3744a', 'width': 1706}, 'variants': {}}]} |
|
What is your use-case for self-hosting an LLM instead of using an API from a provider? | 1 | [removed] | 2025-06-28T10:59:44 | https://www.reddit.com/r/LocalLLaMA/comments/1lmk31o/what_is_your_usecase_for_selfhosting_an_llm/ | g15mouse | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmk31o | false | null | t3_1lmk31o | /r/LocalLLaMA/comments/1lmk31o/what_is_your_usecase_for_selfhosting_an_llm/ | false | false | self | 1 | null |
Benchmarking LLM Inference Libraries for Token Speed & Energy Efficiency | 0 | We conducted a benchmark comparing four popular LLM inference libraries—TensorRT-LLM, vLLM, Ollama, and MLC—in terms of energy per token and tokens per second, using a standardized Docker setup and energy monitoring tools. The benchmark project was original done for a university report.
Experiment Details
• Model: Quantized Qwen2.5-Coder-14B-Instruct
• Different quantized formats were used per library for compatibility: Q4_K_L for Ollama, 4-bit AWQ for vLLM/TensorRT, q4f16_1 for MLC
• Dataset: 80 prompts sampled from the SWE-bench benchmark — real-world GitHub issues
• Each prompt includes issue text + repo context; average length: ~600–700 tokens
• Inference config:
• Max output: 4096 tokens
• Context window: 32k tokens
• Temperature: 0.0 (fully deterministic), Top-K: 20, Top-P: 0.8
• Hardware: NVIDIA RTX 4090 GPU, AMD Ryzen 9 7900X, 64GB RAM
• Energy Measurement: EnergiBridge, logging GPU power and throughput
• Setup: Each inference engine ran 10x inside an identical Docker container, with 60s cooldowns between runs
Notes
• Different quantization per library due to format incompatibilities
• Only tested on NVIDIA GPUs (TensorRT doesn’t support AMD)
• Inference was single-prompt (batch size = 1) due to VRAM constraints as we only had access to 1 GPU
Let me know if there any questions, I originally also wanted to test LMDeploy and SGLang but because of time constraints it was not possible. | 2025-06-28T11:31:47 | https://www.reddit.com/gallery/1lmkmkn | alexbaas3 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1lmkmkn | false | null | t3_1lmkmkn | /r/LocalLLaMA/comments/1lmkmkn/benchmarking_llm_inference_libraries_for_token/ | false | false | 0 | {'enabled': True, 'images': [{'id': 'QJniZOTcXtSaa9bcNvaP-NS4E_BJMhxENtBTX3I0wk4', 'resolutions': [{'height': 64, 'url': 'https://external-preview.redd.it/QJniZOTcXtSaa9bcNvaP-NS4E_BJMhxENtBTX3I0wk4.jpeg?width=108&crop=smart&auto=webp&s=8348840b607d1953ae51352d80841a355ec795d0', 'width': 108}, {'height': 129, 'url': 'https://external-preview.redd.it/QJniZOTcXtSaa9bcNvaP-NS4E_BJMhxENtBTX3I0wk4.jpeg?width=216&crop=smart&auto=webp&s=15b05930f80e8b26cd02a4074c8ecb5522a5e0bc', 'width': 216}, {'height': 192, 'url': 'https://external-preview.redd.it/QJniZOTcXtSaa9bcNvaP-NS4E_BJMhxENtBTX3I0wk4.jpeg?width=320&crop=smart&auto=webp&s=342234d4c88b1021bc31eaf30bda042ed70074af', 'width': 320}, {'height': 384, 'url': 'https://external-preview.redd.it/QJniZOTcXtSaa9bcNvaP-NS4E_BJMhxENtBTX3I0wk4.jpeg?width=640&crop=smart&auto=webp&s=257995fb9672451e0f8e5df831867b9f28ca1db6', 'width': 640}, {'height': 576, 'url': 'https://external-preview.redd.it/QJniZOTcXtSaa9bcNvaP-NS4E_BJMhxENtBTX3I0wk4.jpeg?width=960&crop=smart&auto=webp&s=dd16d606ee6e6fdc5be6f2ec315c829bb96ba6c1', 'width': 960}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/QJniZOTcXtSaa9bcNvaP-NS4E_BJMhxENtBTX3I0wk4.jpeg?auto=webp&s=f5ac0cadc19a279a09d24713c368bbf7472b31a7', 'width': 1000}, 'variants': {}}]} |
|
Using local models with Void | 8 | TLDR; local models like Gemma 27b, Qwen 3 32b can't use the file edit tool in void code
I'm trying to create a simple snake game to test. So far, I've been failing with almost all of the Gemma 4/12/27 models; Qwen 32b seems to do a bit better, but still breaks with editing files.
Anyone has had any luck with Void Code or something similar where these model can use tools correctly? Specifically, I notice that every tool breaks when trying to update the file with 'edit\_file' tool.
LLMs via APIs work perfectly -- which is now starting to give me a feeling that a local setup might not work for even simpler use case
Prompt:
Create a snake game using html and javascript
If you've had better luck, please help | 2025-06-28T12:02:58 | https://www.reddit.com/r/LocalLLaMA/comments/1lml6eo/using_local_models_with_void/ | nuketro0p3r | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lml6eo | false | null | t3_1lml6eo | /r/LocalLLaMA/comments/1lml6eo/using_local_models_with_void/ | false | false | self | 8 | null |
120 AI Chat - Native macOS Chat App with Ollama Support | 0 | Hi everyone,
Just wanted to share a new version of **120 AI Chat**, a native macOS app we've been building that now fully supports local LLMs via Ollama.
**Local Model Support (via Ollama)**
* Llama 3.2
* Mistral 7B
* Deepseek R1
**Useful features for local use**
* Full chat parameter controls (context, temp, penalties, top P)
* Message editing, copying, and deletion
* Fast native performance (built without Electron or browser wrappers)
You can try the app for free, no license key required.
If you like it and want to support us early, you can unlock all features for $39 using the discount code.
We’d love for you to try it out and let us know what you think. We're still actively building and improving, and your feedback would mean a lot!
→ [Download 120 AI Chat](https://120.dev/120-ai-chat#download)
Thanks for checking it out! | 2025-06-28T12:06:18 | https://v.redd.it/uh08zboorn9f1 | 120-dev | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1lml8lx | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/uh08zboorn9f1/DASHPlaylist.mpd?a=1753704393%2CNTcxN2MzMTdjNjVkOGQ5NGI1N2I4ODUyMzBmZGZlZGYyYzYzYWQ5ZDAzYTY2ODFlYWQyODRmMzJmNjQxODM2ZQ%3D%3D&v=1&f=sd', 'duration': 96, 'fallback_url': 'https://v.redd.it/uh08zboorn9f1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/uh08zboorn9f1/HLSPlaylist.m3u8?a=1753704393%2CNmExMDkwZGU3Mzc3ZGZkMTFlZTA0NjNiZTE2YjBlNmFhMWZiYzQxYTVhZTI2NTk5MzQwODliMTcwM2FlMTI2Nw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/uh08zboorn9f1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1560}} | t3_1lml8lx | /r/LocalLLaMA/comments/1lml8lx/120_ai_chat_native_macos_chat_app_with_ollama/ | false | false | 0 | {'enabled': False, 'images': [{'id': 'enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS', 'resolutions': [{'height': 74, 'url': 'https://external-preview.redd.it/enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS.png?width=108&crop=smart&format=pjpg&auto=webp&s=5a2073985c06517281775762f5e68d8dfc4c02bf', 'width': 108}, {'height': 149, 'url': 'https://external-preview.redd.it/enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS.png?width=216&crop=smart&format=pjpg&auto=webp&s=dbefc45c75914eaf0e01889320df949db5b2cc09', 'width': 216}, {'height': 221, 'url': 'https://external-preview.redd.it/enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS.png?width=320&crop=smart&format=pjpg&auto=webp&s=d6e08005292b96a193581580307a865c5a221835', 'width': 320}, {'height': 443, 'url': 'https://external-preview.redd.it/enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS.png?width=640&crop=smart&format=pjpg&auto=webp&s=1740d30f50e5017faf7b16e2de92e956f29d061f', 'width': 640}, {'height': 664, 'url': 'https://external-preview.redd.it/enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS.png?width=960&crop=smart&format=pjpg&auto=webp&s=c60c009efe08628542b5d39b4cf1348d9b1bdcab', 'width': 960}, {'height': 747, 'url': 'https://external-preview.redd.it/enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS.png?width=1080&crop=smart&format=pjpg&auto=webp&s=5f18e32243cdc129440652366d515d4d9020f770', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/enZqcmxkb29ybjlmMQ-0kr0dvIqEj1m9cNIUU1HaIqaql-U6zYfTL-_CgtJS.png?format=pjpg&auto=webp&s=d29d4489add100af588116e993f806d25fcdc5f8', 'width': 1560}, 'variants': {}}]} |
|
Consumer hardware landscape for local LLMs June 2025 | 49 | As a follow-up to [this](https://www.reddit.com/r/LocalLLaMA/comments/1lmf42g/which_is_the_best_16gb_nvidia_gpu_with_balanced/), where OP asked for best 16GB GPU "with balanced price and performance".
For models where "model size" \* "user performance requirements" in total require more bandwidth than CPU/system memory can deliver, there is as of June 2025 no cheaper way than RTX 3090 to get to 24-48-72GB of really fast memory. RTX 3090 still offers the best bang for the buck.
Caveats: At least for inferencing. At this point in time. For a sizeable subset of available models "regular" people want to run at this point in time. With what is considered satisfying performance at this point in time. (YMMV. For me it is good enough quality, slightly faster than I can read.)
Also, LLMs have the same effect as sailboats: you always yearn for the next bigger size.
RTX 3090 is not going to remain on top of that list forever. It is not obvious to me what is going to replace it in the hobbyist space in the immediate future.
My take on the common consumer/prosumer hardware currently available for running LLMs locally:
RTX 3090. Only available as second-hand or (possibly not anymore?) a refurb. **Likely a better option than any non-x090-card in the RTX 4000 or RTX 5000 product lines.**
**I**f you already have a 12GB 3060 or whatever, don't hold off playing with LLMs until you have better hardware! But if you plan to buy hardware for the explicit purpose of playing with LLMs, try to get your hands on a 3090. Because when you eventually want to scale up the \*size\* of the memory, you are very likely going to want the additional memory \*bandwidth\* as well. The 3090 can still be resold, the cost of a new 3060 may be challenging to recover.
RTX 4090 does not offer a compelling performance uplift over 3090 for LLM inferencing, and is 2-2.5x the price as a second-hand option. If you already have one, great. Use it.
RTX 5090 is approaching la-la-land in terms of price/performance for hobbyists. But it \*has\* more memory and better performance.
RTX 6000 Blackwell is actually kind of reasonably priced per GB. But at 8-9k+ USD or whatever, it is still way out of reach for most hobbyists/consumers. Beware of power requirements and (still) some software issues/bugs.
Nvidia DGX Spark (Digits) is definitely interesting. But with "only" 128GB memory, it sort of falls in the middle. Not really enough memory for the big models, too expensive for the small models. Clustering is an option, send more money. Availability is still up in the air, I think.
AMD Strix Halo is a hint at what may come with Medusa Halo (2026) and Gorgon Point (2026-2027). I do not think either of these will come close to match the RTX 3090 in memory bandwidth. But maybe we can get one with 256GB memory? (Not with Strix Halo). And with 256GB, medium sized MoE models may become practical for more of us. (Consumers) We'll see what arrives, and how much it will cost.
Apple Silicon kind of already offers what the AMD APUs (eventually) may deliver in terms of memory bandwidth and size, but tied to OSX and the Apple universe. And the famous Apple tax. Software support appears to be decent.
Intel and AMD are already making stuff which rivals Nvidia's hegemony at the (low end of the) GPU consumer market. The software story is developing, apparently in the right direction.
Very high bar for new contenders on the hardware side, I think. No matter who you are, you are likely going to need commitments from one of Samsung, SK Hynix or Micron in order to actually bring stuff to market *at volume*. And unless you can do it at volume, your stuff will be too expensive for consumers. Qualcomm, Mediatek maybe? Or one of the memory manufacturers themselves. And then, you still need software-support. Either for your custom accelerator/GPU in relevant libraries, or in Linux for your complete system.
It is also possible someone comes up with something insanely smart in software to substantially lower the computational and/or bandwidth cost. For example by combining system memory and GPU memory with smart offloading of caches/layers, which is already a thing. (Curious about how DGX Spark will perform in this setup.) Or maybe someone figures out how to compress current models to a third with no quality loss, thereby reducing the need for memory. For example.
Regular people are still short on *affordable* systems holding at least 256GB or more of memory. Threadripper PRO does exist, but the ones with actual memory bandwidth are not affordable. And neither is 256GB of DDR5 DIMMs.
So, my somewhat opinionated perspective. Feel free to let me know what I have missed. | 2025-06-28T13:10:39 | https://www.reddit.com/r/LocalLLaMA/comments/1lmmh3l/consumer_hardware_landscape_for_local_llms_june/ | ethertype | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1lmmh3l | false | null | t3_1lmmh3l | /r/LocalLLaMA/comments/1lmmh3l/consumer_hardware_landscape_for_local_llms_june/ | false | false | self | 49 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.