title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns] | url
stringlengths 0
780
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns] | 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
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
We made an open source testing agent for UI, API, Vision, Accessibility and Security testing | 1 | End-to-end software test automation has traditionally struggled to keep up with development cycles. Every time the engineering team updates the UI or platforms like **Salesforce** or **SAP** release new updates, maintaining test automation frameworks becomes a bottleneck, slowing down delivery. On top of that, most test automation tools are **expensive** and **difficult to maintain**.
That’s why we built an **open-source AI-powered testing agent**—to make end-to-end test automation **faster, smarter, and accessible** for teams of all sizes.
**High level flow:**
Write natural language tests -> Agent runs the test -> Results, screenshots, network logs, and other traces output to the user.
**Installation:**
pip install testzeus-hercules
**Sample test case for visual testing:**
Feature: This feature displays the image validation capabilities of the agent Scenario Outline: Check if the Github button is present in the hero section Given a user is on the URL as https://testzeus.com And the user waits for 3 seconds for the page to load When the user visually looks for a black colored Github button Then the visual validation should be successful
**Architecture:**
We use AG2 as the base plate for running a multi agentic structure. Tools like Playwright or AXE are used in a REACT pattern for browser automation or accessibility analysis respectively.
**Capabilities**:
The agent can take natural language english tests for UI, API, Accessibility, Security, Mobile and Visual testing. And run them autonomously, so that user does not have to write any code or maintain frameworks.
**Comparison:**
Hercules is a simple open source agent for end to end testing, for people who want to achieve insprint automation.
1. There are multiple testing tools (Tricentis, Functionize, Katalon etc) but not so many agents
2. There are a few testing agents (KaneAI) but its not open source.
3. There are agents, but not built specifically for test automation.
On that last note, we have hardened meta prompts to focus on accuracy of the results.
If you like it, give us a star here: [https://github.com/test-zeus-ai/testzeus-hercules/](https://github.com/test-zeus-ai/testzeus-hercules/) | 2025-02-02T12:41:38 | https://www.reddit.com/r/LocalLLaMA/comments/1ifwexg/we_made_an_open_source_testing_agent_for_ui_api/ | Unhappy-Economics-43 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifwexg | false | null | t3_1ifwexg | /r/LocalLLaMA/comments/1ifwexg/we_made_an_open_source_testing_agent_for_ui_api/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'p_BEHJjh-fVPCpiZzFzkLaC_wycgHuFBsj2EEjHVMv4', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/6sEx_6g4wtdkNPveEngHfUXP9-SZB6KdXqqu7e_Ya10.jpg?width=108&crop=smart&auto=webp&s=9558d48886e7adc4983cccef832e0ed5bd469177', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/6sEx_6g4wtdkNPveEngHfUXP9-SZB6KdXqqu7e_Ya10.jpg?width=216&crop=smart&auto=webp&s=5768243377259639f74e30964d26f89e42a9c4e7', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/6sEx_6g4wtdkNPveEngHfUXP9-SZB6KdXqqu7e_Ya10.jpg?width=320&crop=smart&auto=webp&s=0e3e0d242246a636da7565c18f8c8639900db10d', 'width': 320}, {'height': 480, 'url': 'https://external-preview.redd.it/6sEx_6g4wtdkNPveEngHfUXP9-SZB6KdXqqu7e_Ya10.jpg?width=640&crop=smart&auto=webp&s=4d0d3df39b074b7273d0c1c0f5a05a98ecc01228', 'width': 640}, {'height': 720, 'url': 'https://external-preview.redd.it/6sEx_6g4wtdkNPveEngHfUXP9-SZB6KdXqqu7e_Ya10.jpg?width=960&crop=smart&auto=webp&s=bf3c79d38ae6c2fd0f73c39f6625e731719e4738', 'width': 960}], 'source': {'height': 768, 'url': 'https://external-preview.redd.it/6sEx_6g4wtdkNPveEngHfUXP9-SZB6KdXqqu7e_Ya10.jpg?auto=webp&s=e777f28b6d1cfd42234f256143b24f477c445d48', 'width': 1024}, 'variants': {}}]} |
Sing along with thsi angel | 0 | [https://www.youtube.com/watch?v=ziMvGgq8FQc&list=RDziMvGgq8FQc&start\_radio=1](https://www.youtube.com/watch?v=ziMvGgq8FQc&list=RDziMvGgq8FQc&start_radio=1)
USER: Can you create a phonetic translation of this korean song so that i can sing along?
CHAT GIPPITY: Sure, here's a phonetic transcription of the Korean song lyrics so you can sing along:
\[음악\]
Saeng-gagi manh-eun geon mariya
Dangyeonhi
Haeya hal
Ili-eonnangeon uliga jigeum
Pilsun-iya
Anjeonhan yulibyeong-eul
Pinggyelo
Balam-eul gado
Dun geot gatjiman
Gieokna geu nal-e uliga
Jabassdeon geu soneun mariyeo seol-laeboda
Keodalan mid-eum-i damgyeoseo nan
Geumbang useum-eul jwotjiman
Uleum-i nalkka tto gataseo sojunghan geon
Ijena duryeoumi
Nikka mogsoli neolo
I nam-aissdeon ttalteushan gong-gi
Yeojonhi jasin
Eopsjiman-eun yeogi jeogi salajin
Byeol-ui jari a saranghae
\[음악\]
Handong-an-eul kkaenae bol
Akkim-eops-i banjjag-ineun sigan-eun
Jogeumssig yeol-yeojyeogadeorado
\[음악\]
Jinsim-eul sag-e jebalja hana dul
Chu-eogi tteooleumyeon manh-i manh-i geuliwohal
Geoya
Gomawoss-eoyo
Geulaedo
Ijeneun sageon-ui
Jaepyeongseong neomeolo
\[음악\]
Soljighi duryeobgido hajiman
Nolyeog-eun uli-ege jeongdab-i anilaseo
Majimag seonmul-eun
Santteushan
Annyeong jeog-e sal-a
\[음악\]
\[음악\]
Saeroun
Il mosh-aejun swimeul sag-e jemal-ja hana
Dul chu-eogi tteooleumyeon manh-i manh-i geuli
Mo-eul geoya
Gomawoss-eoyo
Geulaedo
Ijeneun sageon-ui
Jaepyeongseong neomeo
\[음악\]
Lo
Isseul geoya
Akkim-eops-i banjjag-ineun sigan-eun
Jogeumssig
Yeol-yeojyeogadeorado neo
Hana
\[음악\]
Dul chu-eogi
Tteooleumyeon manh-i manh-i geuliwohal geoya
Gomawoss-eoyo
Geulaedo
Ijeneun
Sageon-ui jepumseon neomeolo
\[음악\]
\[음악\]
| 2025-02-02T12:55:13 | https://www.reddit.com/r/LocalLLaMA/comments/1ifwmsr/sing_along_with_thsi_angel/ | DaleCooperHS | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifwmsr | false | null | t3_1ifwmsr | /r/LocalLLaMA/comments/1ifwmsr/sing_along_with_thsi_angel/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'adxPxzhV_iOA6vERaXd5j9UnY7-5uN1t7u3cujmq8no', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/bYXhLCqW4VYddWmA4AcQTDX0XKQd5YYNlswsuaVt-Rg.jpg?width=108&crop=smart&auto=webp&s=5becee46063777240518a989ffaee5b325c84b24', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/bYXhLCqW4VYddWmA4AcQTDX0XKQd5YYNlswsuaVt-Rg.jpg?width=216&crop=smart&auto=webp&s=eda9a9eb5542d15e7063aaa23c42f8736cb63660', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/bYXhLCqW4VYddWmA4AcQTDX0XKQd5YYNlswsuaVt-Rg.jpg?width=320&crop=smart&auto=webp&s=009817d2cde66ab556fa3d1bb3f14d689767c12d', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/bYXhLCqW4VYddWmA4AcQTDX0XKQd5YYNlswsuaVt-Rg.jpg?auto=webp&s=af91d1da71a7e1678a696629a7f4e847646f5c4a', 'width': 480}, 'variants': {}}]} |
How to calculate TPS for multi-request processing | 1 | Hi,
I'm planning on building a rig for running llm's that I will use for production workloads - I'd like to provision a specific thoughput and I'm wondering if my assumptions are correct:
- most t/s values reported here focus on having the model take up most of the memory. If I want to process multiple requests in concurrently, I understand that I effectively need to use a multiple of the memory
- so i.e. If I plan to run a 3gb model and want to be able to process 4 requests at the same time (without batching), I'd need to get at least a ~12gb GPU card (or RAM)
- how is this different for MoE models? I know that they use a smaller chunk of the model size but I suppose the principle stays the same?
- additionally, I noticed that multimodal models require lots more memory compared to text models - ex. Qwen2VL 7B barely fits into an RTX 3090 24gb. How can we accurately measure how much memory is required for multimodal models?
| 2025-02-02T13:15:12 | https://www.reddit.com/r/LocalLLaMA/comments/1ifwzb5/how_to_calculate_tps_for_multirequest_processing/ | lmyslinski | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifwzb5 | false | null | t3_1ifwzb5 | /r/LocalLLaMA/comments/1ifwzb5/how_to_calculate_tps_for_multirequest_processing/ | false | false | self | 1 | null |
How to make a model to read math? | 2 | If I have both the Latex source and the pdf that comes from it, I want a model that can read it even if it is full of math. There doesn't seem anything able to do this currently. As a three month project, does anyone have a recommendation for how I could try to build a model to do this!? | 2025-02-02T13:17:27 | https://www.reddit.com/r/LocalLLaMA/comments/1ifx0ru/how_to_make_a_model_to_read_math/ | MrMrsPotts | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifx0ru | false | null | t3_1ifx0ru | /r/LocalLLaMA/comments/1ifx0ru/how_to_make_a_model_to_read_math/ | false | false | self | 2 | null |
What frontend does everyone use for RAG? | 4 | I have a bunch of documents that i want to throw into a LLM and ask questions about. Mostly Excel sheets that people used as graph paper (text with PDF style structure, including diagrams), as well as a number of PDFs. For privacy reasons im hoping to avoid using LLM websites unless my prompts can be anonymized by something like OpenRouter.
The other day i tried to set up Kotaemon for this but it kept running into issues due to my spotty internet and the GUI was clunky. Are there any better RAG-centric frontends that are easy to set up?
I know RAGFlow is technically the most feature complete but i haven't tried it, and when i did a google search there weren't many reviews of it on reddit. | 2025-02-02T13:32:52 | https://www.reddit.com/r/LocalLLaMA/comments/1ifxan3/what_frontend_does_everyone_use_for_rag/ | SinaMegapolis | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifxan3 | false | null | t3_1ifxan3 | /r/LocalLLaMA/comments/1ifxan3/what_frontend_does_everyone_use_for_rag/ | false | false | self | 4 | null |
Return of the R630s: Can I run a local LLM of my old servers ?! | 1 | [removed] | 2025-02-02T13:46:18 | https://www.reddit.com/r/LocalLLaMA/comments/1ifxja7/return_of_the_r630s_can_i_run_a_local_llm_of_my/ | B4st0s | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifxja7 | false | null | t3_1ifxja7 | /r/LocalLLaMA/comments/1ifxja7/return_of_the_r630s_can_i_run_a_local_llm_of_my/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'tPpqj84eoiNyuncIJE04lHTDWCgqAdsYY6mJzFJFFQ4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/czln2vtpNIvmALqxK-3MkgNjW5RfMT6UBICz4RMfxNw.jpg?width=108&crop=smart&auto=webp&s=e985f769e2fee8446891db8b84dcac382c31538c', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/czln2vtpNIvmALqxK-3MkgNjW5RfMT6UBICz4RMfxNw.jpg?width=216&crop=smart&auto=webp&s=b0d5c933021792240ebf24812d210aeca172e924', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/czln2vtpNIvmALqxK-3MkgNjW5RfMT6UBICz4RMfxNw.jpg?width=320&crop=smart&auto=webp&s=7d703af08c4d827b4ebfabe428583f39f1802266', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/czln2vtpNIvmALqxK-3MkgNjW5RfMT6UBICz4RMfxNw.jpg?width=640&crop=smart&auto=webp&s=baf80c7ac628d68d4407116f9b13d31f641c114e', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/czln2vtpNIvmALqxK-3MkgNjW5RfMT6UBICz4RMfxNw.jpg?width=960&crop=smart&auto=webp&s=00465ac05b546cf393a4f9d05f1d6a29f6acd7d3', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/czln2vtpNIvmALqxK-3MkgNjW5RfMT6UBICz4RMfxNw.jpg?width=1080&crop=smart&auto=webp&s=1d1cc983c001743eb71915ebf992a3cfdc46a24e', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/czln2vtpNIvmALqxK-3MkgNjW5RfMT6UBICz4RMfxNw.jpg?auto=webp&s=0b1f4925b2737d586cad733b84ad8918a4985c10', 'width': 1200}, 'variants': {}}]} |
Output truncation with Llama 3.1 8B | 1 | Hi All
I am running Llama3.1 8B with vLLM, and it's giving me improperly truncated / incomplete output.
I have specified max\_tokens parameter, and it does generate output upto max\_tokens, but it doesn't really understand that it has to finish the reply within max\_tokens, it will just keep on generating and just truncate at max\_tokens, doesn't matter if the last sentence is incomplete or not.
Infact even the whole paragraph sometimes seems incomplete, because it's planning the reply in such a way that it is setting up the stage initially and when max\_tokens are hit, it just stops right there.
I have tried to put it in prompt, I have asked it to keep the response in lets say 2 line, but that didn't work either. It has been so frustrating trying to fix this.
Has anyone been able to fix this issue? even hacky solutions?
| 2025-02-02T13:50:34 | https://www.reddit.com/r/LocalLLaMA/comments/1ifxm19/output_truncation_with_llama_31_8b/ | Icy_Occasion_5277 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifxm19 | false | null | t3_1ifxm19 | /r/LocalLLaMA/comments/1ifxm19/output_truncation_with_llama_31_8b/ | false | false | self | 1 | null |
While some users bully Deepsek R1 with their incoherent inputs, ChatGPT is being absolutly bullied by its creator. What have they done with my boy xd | 1 | 2025-02-02T14:03:28 | JustHangingAroundMan | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ifxux4 | false | null | t3_1ifxux4 | /r/LocalLLaMA/comments/1ifxux4/while_some_users_bully_deepsek_r1_with_their/ | false | false | 1 | {'enabled': True, 'images': [{'id': 'O3wIB4Kdln_2R9IDu6A02rbZV11GHGD0iySxeIt9TEw', 'resolutions': [{'height': 44, 'url': 'https://preview.redd.it/zl5j2u6sfqge1.png?width=108&crop=smart&auto=webp&s=1fab5fc856142daae4ca03d1e95b98640dcb86b8', 'width': 108}, {'height': 89, 'url': 'https://preview.redd.it/zl5j2u6sfqge1.png?width=216&crop=smart&auto=webp&s=ee9d3e0ae6e6e0e5674e1cfd4a45197a9951f1f0', 'width': 216}, {'height': 131, 'url': 'https://preview.redd.it/zl5j2u6sfqge1.png?width=320&crop=smart&auto=webp&s=3731c677c72c29fe667766f019ec0ed142e8d9fa', 'width': 320}, {'height': 263, 'url': 'https://preview.redd.it/zl5j2u6sfqge1.png?width=640&crop=smart&auto=webp&s=b1bc77eeff4214a91fa69e10f8c82c9dd630bd5f', 'width': 640}], 'source': {'height': 338, 'url': 'https://preview.redd.it/zl5j2u6sfqge1.png?auto=webp&s=202b7fb91a6deaa695c53d0f275452b854508fd0', 'width': 820}, 'variants': {}}]} |
|||
Building a Lightweight Inference Platform | 4 | Hey community, so recently, I had built a lightweight Llama 3 inference engine that was written in CUDA C with no additional dependencies. All kernels were custom made. However, there was a lot of room for optimization like tensor core wmma, kv caching, paged attention, and mixed precision inference. Another limitation is the fact that it only runs llama 3.
Tbh, the current tok/s is pretty dog water (especially without kv cache), but either way, my goal was to write CUDA kernels and get my hands dirty. Those are still relatively optimized with memory access, vectorization, warp intrinsics and shared memory optimizations.
I’ve recently received some interest about building an inference platform that is tailored towards inferencing LLMs on a production scale with the same lightweight ideology. The focus would be building something that would work with a variety of models that call kernel backends for general matmul, layernorm, etc. Hence these exposed backends could also enable scaled addition of new models and running proprietary models.
I’d like to gauge if there’s an interest for something like this. Tools like llama.cpp, ollama, TensorRT all exist. At least one issue I can think of is TensorRT is heavy and requires pre compilation of models before aggressively fusing operations and running inference.
I’d still like to hear what the community thinks about this. I would really love to begin building something (open source of course) but only if it has a genuine purpose and , more importantly, meets demand that isn’t being fulfilled currently.
A question I also keep asking myself is at what percentage of cuBLAS performance would someone realistically consider lightweight over leftover performance.
I look forward to hearing your thoughts. | 2025-02-02T14:18:46 | https://github.com/abhisheknair10/llama3.cu | Delicious-Ad-3552 | github.com | 1970-01-01T00:00:00 | 0 | {} | 1ify5l1 | false | null | t3_1ify5l1 | /r/LocalLLaMA/comments/1ify5l1/building_a_lightweight_inference_platform/ | false | false | 4 | {'enabled': False, 'images': [{'id': 'fp0Jmd_8IxuXIDq82aT_dGLnZfhV5cNIERMZ9YWVy_I', 'resolutions': [{'height': 61, 'url': 'https://external-preview.redd.it/rDXna22tsK7TMsSCSlb_Rt3RA4VBEGkMQVLuTuvxBUY.jpg?width=108&crop=smart&auto=webp&s=16b990713f9dd3bbc7af3e75729b128115fd64e9', 'width': 108}, {'height': 123, 'url': 'https://external-preview.redd.it/rDXna22tsK7TMsSCSlb_Rt3RA4VBEGkMQVLuTuvxBUY.jpg?width=216&crop=smart&auto=webp&s=cf767de46742f8b8952c84ef122579112043c164', 'width': 216}, {'height': 182, 'url': 'https://external-preview.redd.it/rDXna22tsK7TMsSCSlb_Rt3RA4VBEGkMQVLuTuvxBUY.jpg?width=320&crop=smart&auto=webp&s=81899fa2743300cb6673c21c028865954edc6292', 'width': 320}, {'height': 365, 'url': 'https://external-preview.redd.it/rDXna22tsK7TMsSCSlb_Rt3RA4VBEGkMQVLuTuvxBUY.jpg?width=640&crop=smart&auto=webp&s=f9811f6a371b03cb11b917b38a3897b30d868c8a', 'width': 640}, {'height': 548, 'url': 'https://external-preview.redd.it/rDXna22tsK7TMsSCSlb_Rt3RA4VBEGkMQVLuTuvxBUY.jpg?width=960&crop=smart&auto=webp&s=dec969d581c01405426af2ad9aa44197a2bf0f28', 'width': 960}, {'height': 617, 'url': 'https://external-preview.redd.it/rDXna22tsK7TMsSCSlb_Rt3RA4VBEGkMQVLuTuvxBUY.jpg?width=1080&crop=smart&auto=webp&s=6179653dd6788f2553e1310f390092a440adcc35', 'width': 1080}], 'source': {'height': 1024, 'url': 'https://external-preview.redd.it/rDXna22tsK7TMsSCSlb_Rt3RA4VBEGkMQVLuTuvxBUY.jpg?auto=webp&s=695005260e25af94d590e4372ccc0ac8816eae8b', 'width': 1792}, 'variants': {}}]} |
|
How national security advisors evaluate tech companies | 1,869 | I just realized I should have added tiktok. | 2025-02-02T14:22:33 | Amgadoz | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ify8f9 | false | null | t3_1ify8f9 | /r/LocalLLaMA/comments/1ify8f9/how_national_security_advisors_evaluate_tech/ | false | false | 1,869 | {'enabled': True, 'images': [{'id': '6oBUDeTO6I9wnm_sOjDO_8GYvlmG77JabDvVdXd3nXA', 'resolutions': [{'height': 61, 'url': 'https://preview.redd.it/h1mbjvbdjqge1.jpeg?width=108&crop=smart&auto=webp&s=de5a2d91a1370fcf7f330f8b507409c55111206d', 'width': 108}, {'height': 123, 'url': 'https://preview.redd.it/h1mbjvbdjqge1.jpeg?width=216&crop=smart&auto=webp&s=fdd44bdb73fe90c4d460e9a19a2b773363113036', 'width': 216}, {'height': 182, 'url': 'https://preview.redd.it/h1mbjvbdjqge1.jpeg?width=320&crop=smart&auto=webp&s=330e21fd47000954143fd558e1958869c4ad7a56', 'width': 320}, {'height': 364, 'url': 'https://preview.redd.it/h1mbjvbdjqge1.jpeg?width=640&crop=smart&auto=webp&s=b8a53ed64a3981fb5114d597cba9461646d11938', 'width': 640}], 'source': {'height': 377, 'url': 'https://preview.redd.it/h1mbjvbdjqge1.jpeg?auto=webp&s=279fb9be61104a9bb49ecf0e33038d58977a6574', 'width': 662}, 'variants': {}}]} |
||
A tutor wants to train a LLM with his course content | 1 | [removed] | 2025-02-02T14:29:32 | https://www.reddit.com/r/LocalLLaMA/comments/1ifydhl/a_tutor_wants_to_train_a_llm_with_his_course/ | gosteneonic | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifydhl | false | null | t3_1ifydhl | /r/LocalLLaMA/comments/1ifydhl/a_tutor_wants_to_train_a_llm_with_his_course/ | false | false | self | 1 | null |
Is this setup Possible ? | 3 | I currenly run OpenWeb UI on my PC with my 4080 Super with ollama serving up various models. Would I be able to configure my local OpenWeb UI to use models running serverless in runpod - this way I have the flexibility of local UI and some local models with the ability to run larger models at optimal cost in the cloud (so I dont have to spend a load on local GPU's ?) if this possible are there any good examples / tutorials or things I should be reading ? | 2025-02-02T14:32:41 | https://www.reddit.com/r/LocalLLaMA/comments/1ifyfvg/is_this_setup_possible/ | droomurray | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifyfvg | false | null | t3_1ifyfvg | /r/LocalLLaMA/comments/1ifyfvg/is_this_setup_possible/ | false | false | self | 3 | null |
we're building Sentient - your personal and private AI with agents and memory | 0 | Sentient is your personal and private AI with agents and memory. We've been working on this for the last 3 months and we just released v1.4 of the app.
Sentient is a desktop app for Windows and it has the following features:
1. Fully local AI inference powered by open source model (Llama 3.2 3B)
2. We pull data from your social media profiles (Reddit, X and LinkedIn), along with a personality test based on MBTI to create a knowledge graph about the user - this knowledge graph powers the memories feature.
3. Memories: Sentient uses a custom graph memory pipeline under the hood to maintain a knowledge graph about the user. As the user interacts with Sentient, it automatically updates the graph - adding or deleting nodes based on the facts it has gleaned from the interaction.
4. Tools and Flows: Sentient is integrated with 6 GSuite tools for now - it can combine these tools together to perform complex tasks. For example, it can create a presentation on Google Slides and email a link to that presentation to someone else.
5. Internet Search: Sentient can search the web to get any additional context it requires to answer queries or complete tasks.
You can check out Sentient at [www.existence.technology/sentient](http://www.existence.technology/sentient)
The website also has an FAQ section that answers the questions we usually get from a lot of folks, a release pipeline that outlines some ideas for future features (you can find this on the Downloads page). We've also put up our privacy policy that clearly mentions what Sentient does with your data. We do not collect any data from users other than a username, their email addresses and login history (this helps us track our early users and also measure retention, churn, etc)
I'd love to hear what you guys think about this and also any additional feature ideas that you may have for the app.
[Here's a sample initial graph that Sentient creates](https://preview.redd.it/zjtsjnv7kqge1.png?width=1722&format=png&auto=webp&s=1c7d29ae390817fb09ed149fad4aa098f3bef97f)
[Sentient learns from user interactions](https://preview.redd.it/ptsgjw8ekqge1.png?width=1396&format=png&auto=webp&s=b94f56ecdabef657603e5d4e6e9cf72747e5b22b)
[An example of a tool call](https://preview.redd.it/s4v7w43jkqge1.png?width=689&format=png&auto=webp&s=083357b310b587a4c4eb58b4e0d603a3340a5874)
| 2025-02-02T14:34:07 | https://www.reddit.com/r/LocalLLaMA/comments/1ifyguy/were_building_sentient_your_personal_and_private/ | therealkabeer | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifyguy | false | null | t3_1ifyguy | /r/LocalLLaMA/comments/1ifyguy/were_building_sentient_your_personal_and_private/ | false | false | 0 | null |
|
New to Local LLM, need some advice | 1 | [removed] | 2025-02-02T14:34:11 | https://www.reddit.com/r/LocalLLaMA/comments/1ifygwj/new_to_local_llm_need_some_advice/ | Berzerk_666 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifygwj | false | null | t3_1ifygwj | /r/LocalLLaMA/comments/1ifygwj/new_to_local_llm_need_some_advice/ | false | false | self | 1 | null |
Best STEM MoE general; reasoning model quant that can run in N (128, 256) GBy RAM+CPU? | 1 | I'm wondering what the best STEM related MoE general purpose as well as reasoning type models' quants are currently that can run in N (128, 256) GBy RAM? i.e. assuming that's your RAM limit but you can use whatever model / quantization level that can be run efficiently on CPU+RAM.
e.g. one could run deepseek-v2.5 in ~240GBy, Q4 in ~128 GBy.
Or one could run deepseek-v3 Q2 in ~172 GBy, 1.58B trinary in somewhat less.
Or one could run some other model like 8x22B (wizardlm, mistral) in ~180GBy Q8, maybe Q5 in ~128GBy though those are older models so may not be competitive even though less quantized due to their much smaller model sizes vs. newer SOTA ones.
Similarly one could perhaps run extremely quantized R1 models or some originally smaller reasoning model with a quantization much closer to Q4, Q5, Q8, BF16 depending on what options may exist / materialize (QwQ, Mistral, ...).
And in any model in some quant.+format configurations one could use non-uniform quantization to have some more essential things quantized at less lossy levels and other less essential things quantized with more highly lossy compressions.
I'm interested in STEM / general knowledge / information type stuff wrt. what is 'best' and generally looking for highly improved capabilities
vs. 32B / 72B level models wrt. accuracy / truthfulness / knowledge / ability to synthesize & correlate / summarize complex information in response to diverse topics.
So ideally we're obviously at the point where the ideal case is to have like 512-768GB RAM on a server with good CPU + RAM BW but I'm not there now, though lots of people could probably cobble together 128, 192, 256 GBy RAM to have something that works but slowly but be left with the question
that given the fortunate choice of a few open models / quants one could marginally run, what's the best performance / utility for such a level of system capability now? | 2025-02-02T14:45:15 | https://www.reddit.com/r/LocalLLaMA/comments/1ifyp1n/best_stem_moe_general_reasoning_model_quant_that/ | Calcidiol | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifyp1n | false | null | t3_1ifyp1n | /r/LocalLLaMA/comments/1ifyp1n/best_stem_moe_general_reasoning_model_quant_that/ | false | false | self | 1 | null |
Would open weighting GPT 3.5 be interesting for us at all? | 57 | Hi guys, please explain if the community would care about such legacy model release? Would the same apply to 4o released in like 2 years? | 2025-02-02T14:52:07 | https://www.reddit.com/r/LocalLLaMA/comments/1ifyu5b/would_open_weighting_gpt_35_be_interesting_for_us/ | robertpiosik | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifyu5b | false | null | t3_1ifyu5b | /r/LocalLLaMA/comments/1ifyu5b/would_open_weighting_gpt_35_be_interesting_for_us/ | false | false | self | 57 | null |
Language Model Recommendation Request | 1 | I am looking for a language model to help me write a story, or a potential film script that I am working on. I need it to have the following characteristics:-
\- It needs to be the most intelligent model I can find, which probably means the highest parameter count. I want to discuss elements of a philosophy, which potentially incorporates elements of quantum mechanics, without necessarily sounding like tabloid "quantum mysticism."
\- It needs to be willing to depict at least one scene of graphic group sex. I don't need it to be uncensored in every respect necessarily, but at least in places, what I am writing is very NSFW.
\- I am willing to use OpenRouter, or otherwise run models via a commercial third party, because I do not currently have the hardware to run a truly large parameter model locally.
\- I need a minimal context of 8,000 tokens, and preferably higher. I don't want a scenario where I only talk ***about*** the characters; I want to write a scenario card where I can actually talk ***to*** the characters, and then re-integrate previously written material, both back into a RAG database, and into my story/film script.
Thank you in advance, for any guidance or suggestions that anyone can give me. | 2025-02-02T14:54:14 | https://www.reddit.com/r/LocalLLaMA/comments/1ifyvoy/language_model_recommendation_request/ | petrus4 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifyvoy | false | null | t3_1ifyvoy | /r/LocalLLaMA/comments/1ifyvoy/language_model_recommendation_request/ | false | false | self | 1 | null |
Mistral Small 3 24b is the first model under 70b I’ve seen pass the “apple” test (even using Q4). | 126 | I put all the Deepseek-R1 distills through the “apple” benchmark last week and only 70b passed the “Write 10 sentences that end with the word “apple” “ test, getting all 10 out of10 sentences correct.
I tested a slew of other newer open source models (all the major ones, Qwen, Phi-, Llama, Gemma, Command-R, etc) as well, but no model under 70b has ever managed to succeed in getting all 10 right….until Mistral Small 3 24b came along.
It is the first and only model under 70b parameters that I’ve found that could pass this test. Congrats Mistral Team!! | 2025-02-02T14:59:50 | https://www.reddit.com/r/LocalLLaMA/comments/1ifyzvv/mistral_small_3_24b_is_the_first_model_under_70b/ | Porespellar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifyzvv | false | null | t3_1ifyzvv | /r/LocalLLaMA/comments/1ifyzvv/mistral_small_3_24b_is_the_first_model_under_70b/ | false | false | self | 126 | null |
Help in choosing right tool for help in academic writing. | 1 | [removed] | 2025-02-02T15:02:57 | https://www.reddit.com/r/LocalLLaMA/comments/1ifz2js/help_in_choosing_right_tool_for_help_in_academic/ | Fun_Repeat_3791 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifz2js | false | null | t3_1ifz2js | /r/LocalLLaMA/comments/1ifz2js/help_in_choosing_right_tool_for_help_in_academic/ | false | false | self | 1 | null |
Next note prediction in DAW | 8 | For the past year I've been using code in-line completions offered by things such as github-copilot, and I wonder if there are similar plugin/software for note completions in DAW (garageband, cubase, logic pro etc.)?
So far music services in the market such as Suno ai generates mp3/wav files, which is like a ChatGPT that outputs pdf instead of plaintext. Therefore, while it is a quite and easy way for hobbyists to generate songs I find it quite useless for more serious usecases, such as music producers who not only want the song to be generated, but also in an editable format (midi) so that the piece can be tweaked and actually used in production.
I know there are quite a few models on huggingface that generate music in editable formats like midi, so I imagine to recreate github copilot for music, we just need some sort of plugin that acts a frontend that detects user input and display the completion, the frontend can connect to either apis (if there are any) or locally served large music models. | 2025-02-02T15:03:22 | https://www.reddit.com/r/LocalLLaMA/comments/1ifz2vm/next_note_prediction_in_daw/ | NoLeading4922 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifz2vm | false | null | t3_1ifz2vm | /r/LocalLLaMA/comments/1ifz2vm/next_note_prediction_in_daw/ | false | false | self | 8 | null |
Is Deepseek stuck in 2023? | 0 | These days I tested Deepseek and I am really disappointed.
It only generates false information, also according to Him we are in 2023.
How can everyone say that it is ok?
https://preview.redd.it/ouswmdlirqge1.png?width=970&format=png&auto=webp&s=d17f93e9dd5aaa35e15598fb632597dd5d35902a
| 2025-02-02T15:08:44 | https://www.reddit.com/r/LocalLLaMA/comments/1ifz709/is_deepseek_stuck_in_2023/ | deadly_poison7 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifz709 | false | null | t3_1ifz709 | /r/LocalLLaMA/comments/1ifz709/is_deepseek_stuck_in_2023/ | false | false | 0 | null |
|
Running LLMs directly from SSD | 1 | [removed] | 2025-02-02T15:15:07 | https://www.reddit.com/r/LocalLLaMA/comments/1ifzc17/running_llms_directly_from_ssd/ | AMOVCS | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifzc17 | false | null | t3_1ifzc17 | /r/LocalLLaMA/comments/1ifzc17/running_llms_directly_from_ssd/ | false | false | self | 1 | null |
4060 Ti 16 GB vs 3090 24 GB | 5 | The local price of used 3090 is $550-$700, whilst the price of a new 4060 Ti 16 GB is $475-$550. I’m going to be buying at least 8x GPUs and building a machine around that. I would like to use this machine as a local server for serving LLMs in my household, but besides that I would like to use it for training LLMs (and other ML models). Going by the price and the VRAM size what is a better option? I’m kinda skeptical about buying used 3090s. | 2025-02-02T15:42:40 | https://www.reddit.com/r/LocalLLaMA/comments/1ifzy48/4060_ti_16_gb_vs_3090_24_gb/ | cybran3 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ifzy48 | false | null | t3_1ifzy48 | /r/LocalLLaMA/comments/1ifzy48/4060_ti_16_gb_vs_3090_24_gb/ | false | false | self | 5 | null |
regarding VRAMs and RAMs | 2 | noob here, distilled 1.b,7b models run on pc without requiring gpus (albient very slow) . So why cant we just slap a lot of 16gm RAM stick together in parallel to run big 70b models | 2025-02-02T15:52:03 | https://www.reddit.com/r/LocalLLaMA/comments/1ig05pf/regarding_vrams_and_rams/ | Narutowale | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig05pf | false | null | t3_1ig05pf | /r/LocalLLaMA/comments/1ig05pf/regarding_vrams_and_rams/ | false | false | self | 2 | null |
Beginner Walkthrough to Install LLMs on Windows | 22 | About six months ago I started a concerted effort to revisit my initial skepticism of LLMs and really try to understand how to get value out of them. As I went through my learning curve, I realized that a lot of the content I was reading either presupposed knowledge I didn't have, or was not easy to follow because of guidelines geared towards using Linux or Mac. I've been writing the guide I had when I started, which I keep updating as new development happen and as I explore things further. I hope this can help newcomers, feedback welcome!
[https://www.oliviertravers.com/running-llms-locally-the-getting-started-windows-stack/](https://www.oliviertravers.com/running-llms-locally-the-getting-started-windows-stack/) | 2025-02-02T15:52:10 | https://www.reddit.com/r/LocalLLaMA/comments/1ig05tf/beginner_walkthrough_to_install_llms_on_windows/ | rdmDgnrtd | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig05tf | false | null | t3_1ig05tf | /r/LocalLLaMA/comments/1ig05tf/beginner_walkthrough_to_install_llms_on_windows/ | false | false | self | 22 | null |
Would External GPU (eGPU) perform well for offline inferencing? | 2 | I use a laptop with external monitors, it has no dedicated GPU (Ryzen 7 Pro). Can I buy an eGPU and connect it to the laptop via thunderbolt to run offline LLM inferencing? If yes, then which RTX desktop GPU would be more appropriate performance and cost wise? | 2025-02-02T15:55:59 | https://www.reddit.com/r/LocalLLaMA/comments/1ig08sh/would_external_gpu_egpu_perform_well_for_offline/ | AMGraduate564 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig08sh | false | null | t3_1ig08sh | /r/LocalLLaMA/comments/1ig08sh/would_external_gpu_egpu_perform_well_for_offline/ | false | false | self | 2 | null |
Need help making LLama.cpp and llama-cpp-python with OpenCL or CLBlast on Termux (Android 10) | 1 | [removed] | 2025-02-02T16:00:35 | https://www.reddit.com/r/LocalLLaMA/comments/1ig0cfs/need_help_making_llamacpp_and_llamacpppython_with/ | Repulsive-Price-9943 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig0cfs | false | null | t3_1ig0cfs | /r/LocalLLaMA/comments/1ig0cfs/need_help_making_llamacpp_and_llamacpppython_with/ | false | false | self | 1 | null |
OpenAI Personal Data Aggregation: I Did Not Really Understand How Personal It Is | 0 | 2025-02-02T16:26:21 | Mbando | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig0y2u | false | null | t3_1ig0y2u | /r/LocalLLaMA/comments/1ig0y2u/openai_personal_data_aggregation_i_did_not_really/ | false | false | 0 | {'enabled': True, 'images': [{'id': 'klwwYcJ-oQR1qU8M7bgbaCVsro99Q6GSgV4mkQFqU30', 'resolutions': [{'height': 96, 'url': 'https://preview.redd.it/ppbwxws75rge1.png?width=108&crop=smart&auto=webp&s=2d691f7eae59b2b1d52678f8514a58c1558de1d5', 'width': 108}, {'height': 193, 'url': 'https://preview.redd.it/ppbwxws75rge1.png?width=216&crop=smart&auto=webp&s=7c48c72f8862710118c824450b360d8e7132ba01', 'width': 216}, {'height': 286, 'url': 'https://preview.redd.it/ppbwxws75rge1.png?width=320&crop=smart&auto=webp&s=808a4306100cfe6b89e2b97c0924cae753bb02b7', 'width': 320}, {'height': 573, 'url': 'https://preview.redd.it/ppbwxws75rge1.png?width=640&crop=smart&auto=webp&s=e282ba0f699a2d75d6ad7f6a911bb0593dcf3631', 'width': 640}, {'height': 860, 'url': 'https://preview.redd.it/ppbwxws75rge1.png?width=960&crop=smart&auto=webp&s=5348bca91ff0cab33bec7f8a47bc132a4f70bccd', 'width': 960}, {'height': 968, 'url': 'https://preview.redd.it/ppbwxws75rge1.png?width=1080&crop=smart&auto=webp&s=2491676083b8ff3c3af046413ab31e7384398153', 'width': 1080}], 'source': {'height': 1182, 'url': 'https://preview.redd.it/ppbwxws75rge1.png?auto=webp&s=c35589f2037df949b91fe5d31ae4d584687cb129', 'width': 1318}, 'variants': {}}]} |
|||
local anonymizer | 0 | I would like to use a local LLM with max 30B to analyze documents with personal data and remove the personal data and insert the letter sequence XXX instead. I used LM Studio with Mistral 7B, LLama 3.1. 8B , Gemma 2 9 B, Deepseek R1 distill Qwen 32B. No model manages to delete all personal data, even though I specify specific data? Does anyone have an idea how this can work? It only works locally because the data is sensitive. | 2025-02-02T16:29:38 | https://www.reddit.com/r/LocalLLaMA/comments/1ig10um/local_anonymizer/ | MTBRiderWorld | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig10um | false | null | t3_1ig10um | /r/LocalLLaMA/comments/1ig10um/local_anonymizer/ | false | false | self | 0 | null |
Have I accidentally found a way to extract way more intelligence out of R1-distill-llama-8B? | 1 | [removed] | 2025-02-02T16:32:07 | https://www.reddit.com/r/LocalLLaMA/comments/1ig12zu/have_i_accidentally_found_a_way_to_extract_way/ | Valuable-Run2129 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig12zu | false | null | t3_1ig12zu | /r/LocalLLaMA/comments/1ig12zu/have_i_accidentally_found_a_way_to_extract_way/ | false | false | self | 1 | null |
How can I do this ? (Google LM notebook overview) | 0 | I mean, I can generate the podcast overview on my LM Notebook. But how can I do that with local LLM with the same quality of natural conversation ? What pipelines and models that I can use ? | 2025-02-02T16:34:30 | https://youtu.be/hpve4d1RejE?si=Xsu7Lp38YHEAfPu4 | Boring-Test5522 | youtu.be | 1970-01-01T00:00:00 | 0 | {} | 1ig14wd | false | {'oembed': {'author_name': 'Bamby Media', 'author_url': 'https://www.youtube.com/@bambymedia', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/hpve4d1RejE?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="Google NotebookLM - Creating a 2 person podcast with NO HUMANS"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/hpve4d1RejE/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'Google NotebookLM - Creating a 2 person podcast with NO HUMANS', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'} | t3_1ig14wd | /r/LocalLLaMA/comments/1ig14wd/how_can_i_do_this_google_lm_notebook_overview/ | false | false | 0 | {'enabled': False, 'images': [{'id': '65EN7Obs2gbqcJ1XSZMdqUShV1BiDp-zBWe7deePJEc', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/kFvc1jIkADFOzsJNgjjeaMqADjZPGrIsYFOIImXr_pM.jpg?width=108&crop=smart&auto=webp&s=1a0757dd01ddef161c15fb2e95a8c01aada0be24', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/kFvc1jIkADFOzsJNgjjeaMqADjZPGrIsYFOIImXr_pM.jpg?width=216&crop=smart&auto=webp&s=89f584f7eb88e903429ce76590bfbe42b366991f', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/kFvc1jIkADFOzsJNgjjeaMqADjZPGrIsYFOIImXr_pM.jpg?width=320&crop=smart&auto=webp&s=32333b74015b3bec853d8b9a7ee17b331391ae70', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/kFvc1jIkADFOzsJNgjjeaMqADjZPGrIsYFOIImXr_pM.jpg?auto=webp&s=5e9fb2ca9638bee8340385923056a70f18c2f4ac', 'width': 480}, 'variants': {}}]} |
|
Why Does My Locally-Run DeepSeek R1 Model Give Inconsistent Responses to the Same Prompt? | 1 | [removed] | 2025-02-02T16:41:03 | https://www.reddit.com/r/LocalLLaMA/comments/1ig1a9w/why_does_my_locallyrun_deepseek_r1_model_give/ | Due-Collar2748 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig1a9w | false | null | t3_1ig1a9w | /r/LocalLLaMA/comments/1ig1a9w/why_does_my_locallyrun_deepseek_r1_model_give/ | false | false | self | 1 | null |
Reading on Distributed AI infra | 2 | Can folks recommend any blogs / papers or sources about building distributed systems for AI workloads? I’m looking at find what are good architectures around serving and scaling inference models, training , fine tuning infrastructure. | 2025-02-02T16:59:40 | https://www.reddit.com/r/LocalLLaMA/comments/1ig1q7g/reading_on_distributed_ai_infra/ | coolsank | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig1q7g | false | null | t3_1ig1q7g | /r/LocalLLaMA/comments/1ig1q7g/reading_on_distributed_ai_infra/ | false | false | self | 2 | null |
Dual 3080ti’s or a 3090 | 1 | Another what should I buy post….. I have an under utilized 3080ti already which has virtually the same bandwidth as a 3090 just less capacity (obviously). I can get another 3080ti for around $450 vs $800+ for a 3090. Or if I get a 3090, what about pairing that with my 3080ti? What are the advantages/disadvantages to dual 3080tis? What about a single 3090 or a 3090 + 3080ti? Would love some insight. I have a few options for the rest of the system (a couple of Ryzen desktops with dual 16x slots or an Epyc 7551p but would need to get a motherboard). Would like to add a deepseek instance to my homelab just not sure what makes the most sense. | 2025-02-02T17:06:42 | https://www.reddit.com/r/LocalLLaMA/comments/1ig1wkn/dual_3080tis_or_a_3090/ | paulrbeers | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig1wkn | false | null | t3_1ig1wkn | /r/LocalLLaMA/comments/1ig1wkn/dual_3080tis_or_a_3090/ | false | false | self | 1 | null |
Everyone with deepseek rn | 12 | [removed] | 2025-02-02T17:18:27 | The_GSingh | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig26oa | false | null | t3_1ig26oa | /r/LocalLLaMA/comments/1ig26oa/everyone_with_deepseek_rn/ | false | false | 12 | {'enabled': True, 'images': [{'id': '2IxUb8mlJA4dejbnJCQRN7I26fty6RA3pSvHPPANy6c', 'resolutions': [{'height': 137, 'url': 'https://preview.redd.it/qll6mn5rerge1.jpeg?width=108&crop=smart&auto=webp&s=8515ccfdb956b2cf131814b2a243c5f0d39e21d9', 'width': 108}, {'height': 274, 'url': 'https://preview.redd.it/qll6mn5rerge1.jpeg?width=216&crop=smart&auto=webp&s=9219730460479eb3168feeb54cd8968172ec6b27', 'width': 216}, {'height': 407, 'url': 'https://preview.redd.it/qll6mn5rerge1.jpeg?width=320&crop=smart&auto=webp&s=11699b94eb9cc99e21bed3450d3bef933f48dcab', 'width': 320}, {'height': 814, 'url': 'https://preview.redd.it/qll6mn5rerge1.jpeg?width=640&crop=smart&auto=webp&s=2459d9cef1839f0840d3a63a86456022982443f5', 'width': 640}, {'height': 1221, 'url': 'https://preview.redd.it/qll6mn5rerge1.jpeg?width=960&crop=smart&auto=webp&s=cb7c021d071295fa3511cd1e100f053479aafa58', 'width': 960}, {'height': 1374, 'url': 'https://preview.redd.it/qll6mn5rerge1.jpeg?width=1080&crop=smart&auto=webp&s=fec55953310df81b7faf3dac6baf7eb04aab53ce', 'width': 1080}], 'source': {'height': 1500, 'url': 'https://preview.redd.it/qll6mn5rerge1.jpeg?auto=webp&s=fae52e6d4c88a98e70f6b26c2881d32f3f62f784', 'width': 1179}, 'variants': {}}]} |
||
gpu problem Lm studio - I just downloaded it, and it doesn't see my graphics card, I'm not sure what the problem is, I can't even launch the starter model. | 1 | [removed] | 2025-02-02T17:20:37 | [deleted] | 1970-01-01T00:00:00 | 0 | {} | 1ig28iz | false | null | t3_1ig28iz | /r/LocalLLaMA/comments/1ig28iz/gpu_problem_lm_studio_i_just_downloaded_it_and_it/ | false | false | default | 1 | null |
||
mistral-small-24b-instruct-2501 is simply the best model ever made. | 967 | It’s the only truly good model that can run locally on a normal machine. I'm running it on my M3 36GB and it performs fantastically with 18 TPS (tokens per second). It responds to everything precisely for day-to-day use, serving me as well as ChatGPT does.
For the first time, I see a local model actually delivering satisfactory results. Does anyone else think so? | 2025-02-02T17:25:29 | https://www.reddit.com/r/LocalLLaMA/comments/1ig2cm2/mistralsmall24binstruct2501_is_simply_the_best/ | hannibal27 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig2cm2 | false | null | t3_1ig2cm2 | /r/LocalLLaMA/comments/1ig2cm2/mistralsmall24binstruct2501_is_simply_the_best/ | false | false | self | 967 | null |
What are people using Deepseek R1 for? | 0 | I know that R1 is used a lot for coding tasks, but what other use cases is everyone using it for? | 2025-02-02T17:42:04 | https://www.reddit.com/r/LocalLLaMA/comments/1ig2qxo/what_are_people_using_deepseek_r1_for/ | Born_Search2534 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig2qxo | false | null | t3_1ig2qxo | /r/LocalLLaMA/comments/1ig2qxo/what_are_people_using_deepseek_r1_for/ | false | false | self | 0 | null |
Feedback on my build for video editing and distilled R1 model: where to invest or cut? | 1 | [removed] | 2025-02-02T17:48:26 | https://www.reddit.com/r/LocalLLaMA/comments/1ig2win/feedback_on_my_build_for_video_editing_and/ | ulisse1988 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig2win | false | null | t3_1ig2win | /r/LocalLLaMA/comments/1ig2win/feedback_on_my_build_for_video_editing_and/ | false | false | self | 1 | null |
Absolute Best Hardware | 1 | Within a $1000 or so budget.
Best hardware path to go down for local hosting a range of the most powerful/large models?
I have a lot going on right now and spending days researching this is giving me dread just thinking about it.
Say I had about $1000 Canadian to spend? Will I need much more? I want to run the biggest and best (practically speaking) models
I do have an older dell xps workstation laptop with an Nvidia chip, wonder if I could just repurpose that? I think it's like a 7th or 8th gen Intel with a workstation Nvidia dedicated laptop graphics card.. | 2025-02-02T17:59:52 | https://www.reddit.com/r/LocalLLaMA/comments/1ig36cr/absolute_best_hardware/ | TraditionalAd552 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig36cr | false | null | t3_1ig36cr | /r/LocalLLaMA/comments/1ig36cr/absolute_best_hardware/ | false | false | self | 1 | null |
How to get attention maps from vision encoder of multimodal LLMs like Llama-3.2-Vision? | 2 | I am working on a project that will allow the user to see what the model "sees" of the input image during each token generation. I want to extract attention maps from the image encoder of the model during inference. Is there any code for this or any idea how to do this? | 2025-02-02T18:06:08 | https://www.reddit.com/r/LocalLLaMA/comments/1ig3c2i/how_to_get_attention_maps_from_vision_encoder_of/ | SussyAmogusChungus | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig3c2i | false | null | t3_1ig3c2i | /r/LocalLLaMA/comments/1ig3c2i/how_to_get_attention_maps_from_vision_encoder_of/ | false | false | self | 2 | null |
🔥 Chipper RAG Toolbox 2.2 is Here! (Ollama API Reflection, DeepSeek, Haystack, Python) | 3 | Big news for all **Ollama** and **RAG** enthusiasts – **Chipper 2.2** is out, and it's packing some serious upgrades!
**Chipper Chains,** you can now link multiple Chipper instances together, distributing workloads across servers and pushing the ultimate context boundary. Just set your `OLLAMA_URL` to another Chipper instance, and lets go.
💡 **What's new?**
\- **Full Ollama API Reflection** – Chipper is now a seamless drop-in service that fully mirrors the **Ollama Chat API**, integrating **RAG capabilities** without breaking existing workflows.
\- **API Proxy & Security** – Reflects & proxies **non-RAG pipeline calls**, with **bearer token support** for a **more secure** Ollama setup.
\- **Daisy-Chaining** – Connect multiple **Chipper** instances to extend processing across multiple nodes.
\- **Middleware** – Chipper now acts as an **Ollama middleware**, also enabling **client-side query parameters** for fine-tuned responses or server side overrides.
\- **DeepSeek R1 Support** \- The Chipper web UI does now supports <think> tags.
⚡ **Why this matters?**
* Easily add **shared RAG capabilities** to your favourite **Ollama Client** with **little extra complexity**.
* Securely expose your **Ollama** server to desktop clients (like **Enchanted**) with bearer token support.
* Run multi-instance **RAG pipelines** to augment requests with distributed knowledge bases or services.
If you find Chipper useful or exciting, **leaving a star would be lovely** and will help others discover Chipper too ✨. I am working on many more ideas and occasionally want to share my progress here with you.
For everyone upgrading to version 2.2, please regenerate your `.env` files using the `run` tool, and don't forget to regenerate your images.
🔗 **Check it out & demo it yourself:**
👉 [https://github.com/TilmanGriesel/chipper](https://github.com/TilmanGriesel/chipper)
👉 [https://chipper.tilmangriesel.com/](https://chipper.tilmangriesel.com/)
**Get started:** [https://chipper.tilmangriesel.com/get-started.html](https://chipper.tilmangriesel.com/get-started.html)
https://i.redd.it/y8kq2y36lrge1.gif
https://i.redd.it/6j46hz77lrge1.gif
https://i.redd.it/o9cfokr7lrge1.gif
| 2025-02-02T18:06:08 | https://www.reddit.com/r/LocalLLaMA/comments/1ig3c2k/chipper_rag_toolbox_22_is_here_ollama_api/ | Alarming_Divide_1339 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig3c2k | false | null | t3_1ig3c2k | /r/LocalLLaMA/comments/1ig3c2k/chipper_rag_toolbox_22_is_here_ollama_api/ | false | false | 3 | {'enabled': False, 'images': [{'id': 'E-kEI5197G8kM3X9w_elSmIqON0oxPvPncXwN6ZG2MY', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/EDUvohhVb5xr-KRAeuTg8gg3QUUyDGnrLD58QPihBNs.jpg?width=108&crop=smart&auto=webp&s=dab89f0ef98812ab4e7c857b48cb84ac2407042b', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/EDUvohhVb5xr-KRAeuTg8gg3QUUyDGnrLD58QPihBNs.jpg?width=216&crop=smart&auto=webp&s=c148d2ee49345abe0b778295f8d0db40061a6406', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/EDUvohhVb5xr-KRAeuTg8gg3QUUyDGnrLD58QPihBNs.jpg?width=320&crop=smart&auto=webp&s=36c1db5ac831e1cab6bec08c030c004aa0068232', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/EDUvohhVb5xr-KRAeuTg8gg3QUUyDGnrLD58QPihBNs.jpg?width=640&crop=smart&auto=webp&s=811d33bdc4c31e21572a972d29ec5913e016dd45', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/EDUvohhVb5xr-KRAeuTg8gg3QUUyDGnrLD58QPihBNs.jpg?width=960&crop=smart&auto=webp&s=944b48d019df668b23a18cc8588c864466577d77', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/EDUvohhVb5xr-KRAeuTg8gg3QUUyDGnrLD58QPihBNs.jpg?width=1080&crop=smart&auto=webp&s=01e0597e135d6c5755efe8d6b609451aa6d36f2e', 'width': 1080}], 'source': {'height': 640, 'url': 'https://external-preview.redd.it/EDUvohhVb5xr-KRAeuTg8gg3QUUyDGnrLD58QPihBNs.jpg?auto=webp&s=8f8d8a351a71526f6d8ee8f233be68854b9d6fb2', 'width': 1280}, 'variants': {}}]} |
|
Mistral small 3 through Openrouter is broken, while it works great with the exact same prompts through other providers and the official API | 25 | 2025-02-02T18:17:46 | HIVVIH | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig3m32 | false | null | t3_1ig3m32 | /r/LocalLLaMA/comments/1ig3m32/mistral_small_3_through_openrouter_is_broken/ | false | false | 25 | {'enabled': True, 'images': [{'id': '7pD3Gk6SDFCujzbATW1ji6F7dWLHxjWRsHNyZt4En1s', 'resolutions': [{'height': 95, 'url': 'https://preview.redd.it/mwax9fnbprge1.png?width=108&crop=smart&auto=webp&s=99cb07626dcfbb44179b2391c6ba4421a4384c21', 'width': 108}, {'height': 191, 'url': 'https://preview.redd.it/mwax9fnbprge1.png?width=216&crop=smart&auto=webp&s=f7eeb6b812b51bd6765aabe9610586c03e4730e9', 'width': 216}, {'height': 283, 'url': 'https://preview.redd.it/mwax9fnbprge1.png?width=320&crop=smart&auto=webp&s=517da62a209f78f4de1794a89e621d14497a43bc', 'width': 320}, {'height': 566, 'url': 'https://preview.redd.it/mwax9fnbprge1.png?width=640&crop=smart&auto=webp&s=7e20eb50542a043d90fbbd09b00d1b8f8ffea5e4', 'width': 640}, {'height': 850, 'url': 'https://preview.redd.it/mwax9fnbprge1.png?width=960&crop=smart&auto=webp&s=88ec95694b4c1a740a1c4ec979ccf9a1d078f755', 'width': 960}, {'height': 956, 'url': 'https://preview.redd.it/mwax9fnbprge1.png?width=1080&crop=smart&auto=webp&s=8613669a91efc579f3c38a4c1663956f89f5e83a', 'width': 1080}], 'source': {'height': 1255, 'url': 'https://preview.redd.it/mwax9fnbprge1.png?auto=webp&s=b1fce39024bbefed65800df683952cb16e3f154e', 'width': 1417}, 'variants': {}}]} |
|||
Qwen Chat Major Update: Qwen2.5-Plus (Closer to Max), Flexible Modes & Unlimited Inputs | 16 | 2025-02-02T18:20:22 | McSnoo | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig3ob4 | false | null | t3_1ig3ob4 | /r/LocalLLaMA/comments/1ig3ob4/qwen_chat_major_update_qwen25plus_closer_to_max/ | false | false | 16 | {'enabled': True, 'images': [{'id': 'uX9pCSUHDRIhBUGKMNM7M3NrZUWReLNTEzuzOGJqdAo', 'resolutions': [{'height': 175, 'url': 'https://preview.redd.it/8z6tl9nsprge1.jpeg?width=108&crop=smart&auto=webp&s=0962a8b7b84123d054543ad1da3ad39fcc075db9', 'width': 108}, {'height': 350, 'url': 'https://preview.redd.it/8z6tl9nsprge1.jpeg?width=216&crop=smart&auto=webp&s=1ccded933c384dc94c24576223f49eea00a8a472', 'width': 216}, {'height': 518, 'url': 'https://preview.redd.it/8z6tl9nsprge1.jpeg?width=320&crop=smart&auto=webp&s=461486ff600cffb7e98f1542b04f489d5b2393b0', 'width': 320}, {'height': 1037, 'url': 'https://preview.redd.it/8z6tl9nsprge1.jpeg?width=640&crop=smart&auto=webp&s=cdf9fc07b664eee4cf217a7062441414af875a28', 'width': 640}, {'height': 1556, 'url': 'https://preview.redd.it/8z6tl9nsprge1.jpeg?width=960&crop=smart&auto=webp&s=17825c37ffe1c4b20daf1660c5469b8495d41d6c', 'width': 960}, {'height': 1751, 'url': 'https://preview.redd.it/8z6tl9nsprge1.jpeg?width=1080&crop=smart&auto=webp&s=512b8e1ede77a74855d5de6c28aef060afcb808a', 'width': 1080}], 'source': {'height': 1751, 'url': 'https://preview.redd.it/8z6tl9nsprge1.jpeg?auto=webp&s=03bcdd23e05ab117039155767db1f7376d743ccb', 'width': 1080}, 'variants': {}}]} |
|||
Is AWQ 8-bit quantization a thing? | 5 | Basically the title.
I saw that GPTQ quantization had an 8-bit version but I was wondering whether people actually use AWQ in 8-bit. Apparently the AutoAWQ library has 8-bit as an option but do inference engines like vLLM support it?
I haven't seen any AWQ 8-bit quantizations to this day so I was just wondering. | 2025-02-02T18:35:07 | https://www.reddit.com/r/LocalLLaMA/comments/1ig41d3/is_awq_8bit_quantization_a_thing/ | random-tomato | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig41d3 | false | null | t3_1ig41d3 | /r/LocalLLaMA/comments/1ig41d3/is_awq_8bit_quantization_a_thing/ | false | false | self | 5 | null |
Qwen/deepseek in colab | 1 | [removed] | 2025-02-02T18:41:27 | https://www.reddit.com/r/LocalLLaMA/comments/1ig470r/qwendeepseek_in_colab/ | Agreeable-Cod-1487 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig470r | false | null | t3_1ig470r | /r/LocalLLaMA/comments/1ig470r/qwendeepseek_in_colab/ | false | false | self | 1 | null |
Medtator, local llama medical research helper | 1 | [deleted] | 2025-02-02T18:50:31 | [deleted] | 1970-01-01T00:00:00 | 0 | {} | 1ig4evs | false | null | t3_1ig4evs | /r/LocalLLaMA/comments/1ig4evs/medtator_local_llama_medical_research_helper/ | false | false | default | 1 | null |
||
Medtator, local Llama research medical helper | 53 | Built a Medical AI Chatbot Using Llama 3 and PubMed
Hey everyone, I’ve been working on a project that combines Llama 3 with PubMed to create an AI-powered medical research assistant. Meet Medtator—a chatbot that retrieves, summarizes, and explains medical literature from PubMed in an accessible way.
How It Works:
- User asks a medical question in natural language.
- Keyword extraction (via a transformer-based pipeline) identifies relevant biomedical terms.
- PubTator API fetches PubMed articles, ranking the most relevant studies.
- Llama 3 generates concise answers using the retrieved context.
Why?
Searching PubMed is a pain. If you’re a medical student, researcher, or just a nerd like me, you know how frustrating it is to sift through thousands of articles. Medtator helps find relevant papers quickly and provides AI-assisted summaries to save time.
Challenges & Learnings:
• Keyword extraction is tricky—LLMs sometimes miss the right biomedical terms.
• Context window limitations force careful article selection to avoid hallucination.
• Fine-tuning vs. prompt engineering—so far, smart prompting works surprisingly well.
Still improving it, but I’d love to hear your thoughts on optimizing Llama 3 for retrieval-augmented generation (RAG) tasks. Would love to hear any feedback/recommendations:). | 2025-02-02T18:51:52 | RaiRamz | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig4g1u | false | null | t3_1ig4g1u | /r/LocalLLaMA/comments/1ig4g1u/medtator_local_llama_research_medical_helper/ | false | false | 53 | {'enabled': True, 'images': [{'id': 'wERAxIx7pyI_NGekFc7cAekFJuWpsvuuFd_DiA-ADxY', 'resolutions': [{'height': 51, 'url': 'https://preview.redd.it/g7qcnt3fvrge1.jpeg?width=108&crop=smart&auto=webp&s=ab5258c6d71825ed4278eca0337996b60f6b28f4', 'width': 108}, {'height': 103, 'url': 'https://preview.redd.it/g7qcnt3fvrge1.jpeg?width=216&crop=smart&auto=webp&s=1f901655c9f154731908de0dbb7b9e7827797a0e', 'width': 216}, {'height': 153, 'url': 'https://preview.redd.it/g7qcnt3fvrge1.jpeg?width=320&crop=smart&auto=webp&s=d7b15fe0c6b0916df421973378845ad61452d2d1', 'width': 320}, {'height': 306, 'url': 'https://preview.redd.it/g7qcnt3fvrge1.jpeg?width=640&crop=smart&auto=webp&s=a0da08cfc3f47f01ce1481a781c7b65ab51fcc7f', 'width': 640}, {'height': 459, 'url': 'https://preview.redd.it/g7qcnt3fvrge1.jpeg?width=960&crop=smart&auto=webp&s=b1744d33cf5c589eb31d0c7bb8b532b0afbe4b81', 'width': 960}, {'height': 516, 'url': 'https://preview.redd.it/g7qcnt3fvrge1.jpeg?width=1080&crop=smart&auto=webp&s=caac8526f330fa6d5f0fc128a52ce7d9fa663f36', 'width': 1080}], 'source': {'height': 912, 'url': 'https://preview.redd.it/g7qcnt3fvrge1.jpeg?auto=webp&s=0826163d5f2612095ef0f3999fff5fbfd4bb745b', 'width': 1906}, 'variants': {}}]} |
||
Please help me identify this TTS on this YouTube video | 1 | [removed] | 2025-02-02T19:01:21 | https://www.reddit.com/r/LocalLLaMA/comments/1ig4odm/please_help_me_identify_this_tts_on_this_youtube/ | redchrism | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig4odm | false | null | t3_1ig4odm | /r/LocalLLaMA/comments/1ig4odm/please_help_me_identify_this_tts_on_this_youtube/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': '_QaApVwuSvlARhU5fbRYCpkchAhMrjbuS3rwts0mLY8', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/0RU_Z-xAFP2510nY5wY5AAAWnIt76OYvVeCSszQjBjE.jpg?width=108&crop=smart&auto=webp&s=3855a3f6c0ee8f3dc9a23a67137f2f7b8e7a3946', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/0RU_Z-xAFP2510nY5wY5AAAWnIt76OYvVeCSszQjBjE.jpg?width=216&crop=smart&auto=webp&s=e30394202690ff2fc5b5c0d1a14677f43457f4a4', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/0RU_Z-xAFP2510nY5wY5AAAWnIt76OYvVeCSszQjBjE.jpg?width=320&crop=smart&auto=webp&s=df144cf4d5873ae3ec632bfb73908961ac8493e3', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/0RU_Z-xAFP2510nY5wY5AAAWnIt76OYvVeCSszQjBjE.jpg?auto=webp&s=8b8587f8013cf65e0d2fe104b762ec9754ee7c2f', 'width': 480}, 'variants': {}}]} |
Local Quantization Workflows: what tools do you use? | 1 | [removed] | 2025-02-02T19:05:03 | https://www.reddit.com/r/LocalLLaMA/comments/1ig4rrk/local_quantization_workflows_what_tools_do_you_use/ | SilentChip5913 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig4rrk | false | null | t3_1ig4rrk | /r/LocalLLaMA/comments/1ig4rrk/local_quantization_workflows_what_tools_do_you_use/ | false | false | self | 1 | null |
I asked DeepSeek R1 and couldn't answer my question | 1 | 2025-02-02T19:07:18 | Bassiette03 | i.ibb.co | 1970-01-01T00:00:00 | 0 | {} | 1ig4tqp | false | null | t3_1ig4tqp | /r/LocalLLaMA/comments/1ig4tqp/i_asked_deepseek_r1_and_couldnt_answer_my_question/ | false | false | 1 | {'enabled': True, 'images': [{'id': 'DnNziwjdeFEKJfLUbINZB6v0NujT3LX9sbHzopdxTSE', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/kxrVpoDEo7EmMYwzCzyI8i2FkI9WRcbESPt0nTWDGgM.jpg?width=108&crop=smart&auto=webp&s=9b4745f7488208c06f0a4ae3d036cb1d4f136e7f', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/kxrVpoDEo7EmMYwzCzyI8i2FkI9WRcbESPt0nTWDGgM.jpg?width=216&crop=smart&auto=webp&s=f3e3f39425b67939f9fa751b912e65e83eb5bc3e', 'width': 216}, {'height': 640, 'url': 'https://external-preview.redd.it/kxrVpoDEo7EmMYwzCzyI8i2FkI9WRcbESPt0nTWDGgM.jpg?width=320&crop=smart&auto=webp&s=2850194e1a73ae66951ed98d44fd61999c87fbc3', 'width': 320}, {'height': 1280, 'url': 'https://external-preview.redd.it/kxrVpoDEo7EmMYwzCzyI8i2FkI9WRcbESPt0nTWDGgM.jpg?width=640&crop=smart&auto=webp&s=614a7345ece708193a980c32e47949ae89506406', 'width': 640}, {'height': 1920, 'url': 'https://external-preview.redd.it/kxrVpoDEo7EmMYwzCzyI8i2FkI9WRcbESPt0nTWDGgM.jpg?width=960&crop=smart&auto=webp&s=76e312ed5156a0da17bf070829835f6b962bf48c', 'width': 960}, {'height': 2160, 'url': 'https://external-preview.redd.it/kxrVpoDEo7EmMYwzCzyI8i2FkI9WRcbESPt0nTWDGgM.jpg?width=1080&crop=smart&auto=webp&s=456f8a0c5391b0849f4462667f3aaf917167f446', 'width': 1080}], 'source': {'height': 2412, 'url': 'https://external-preview.redd.it/kxrVpoDEo7EmMYwzCzyI8i2FkI9WRcbESPt0nTWDGgM.jpg?auto=webp&s=3f57f4a01854a2c522b2708ac04e047fc3ec097d', 'width': 1080}, 'variants': {}}]} |
|||
Which DeepSeek model is the "funny" one? | 1 | [removed] | 2025-02-02T19:40:37 | https://www.reddit.com/r/LocalLLaMA/comments/1ig5mp6/which_deepseek_model_is_the_funny_one/ | Accomplished_Smell53 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig5mp6 | false | null | t3_1ig5mp6 | /r/LocalLLaMA/comments/1ig5mp6/which_deepseek_model_is_the_funny_one/ | false | false | self | 1 | null |
Can you run DeepSeek R1 distilled version with web search? | 0 | I was using LM studio because it's easy for a quick install (I have ollama too but have not set that one up yet).
In LM studio, I do not see a way to run any of the DeepSeek R1 models with web search. Am I missing something? Do I need to run R1 through Ollama instead to get web search? | 2025-02-02T19:41:53 | https://www.reddit.com/r/LocalLLaMA/comments/1ig5nsb/can_you_run_deepseek_r1_distilled_version_with/ | Design-Dragon | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig5nsb | false | null | t3_1ig5nsb | /r/LocalLLaMA/comments/1ig5nsb/can_you_run_deepseek_r1_distilled_version_with/ | false | false | self | 0 | null |
Hardware recommendation for running 70b AMP (up to FP32 prob) | 1 | [removed] | 2025-02-02T19:49:29 | https://www.reddit.com/r/LocalLLaMA/comments/1ig5u7r/hardware_recommendation_for_running_70b_amp_up_to/ | zekken523 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig5u7r | false | null | t3_1ig5u7r | /r/LocalLLaMA/comments/1ig5u7r/hardware_recommendation_for_running_70b_amp_up_to/ | false | false | self | 1 | null |
New Docker Guide for R2R's (Reason-to-Retrieve) local AI system | 8 | Hey r/LocalLLaMA,
I just put together a quick beginner’s guide for [**R2R**](https://r2r-docs.sciphi.ai/introduction) — an all-in-one open source AI Retrieval-Augmented Generation system that’s easy to self-host and super flexible for a range of use cases. R2R lets you ingest documents (PDFs, images, audio, JSON, etc.) into a local or cloud-based knowledge store, and then query them using advanced hybrid or graph-based search. It even supports multi-step “agentic” reasoning if you want more powerful question answering, coding hints, or domain-specific Q&A on your private data.
I’ve included some references and commands below for anyone new to Docker or Docker Swarm. If you have any questions, feel free to ask!
# Link-List
|Service|Link|
|:-|:-|
|**Owners Website**|[https://sciphi.ai/](https://sciphi.ai/)|
|**GitHub**|[https://github.com/SciPhi-AI/R2R](https://github.com/SciPhi-AI/R2R)|
|**Docker & Full Installation Guide**|[Self-Hosting (Docker)](https://r2r-docs.sciphi.ai/self-hosting/installation/full/docker)|
|**Quickstart Docs**|[R2R Quickstart](https://r2r-docs.sciphi.ai/self-hosting/quickstart)|
# Basic Setup Snippet
**1. Install the CLI & Python SDK -**
pip install r2r
**2. Launch R2R with Docker**(This command pulls all necessary images and starts the R2R stack — including Postgres/pgvector and the Hatchet ingestion service.)
export OPENAI\_API\_KEY=sk-...
r2r serve --docker --full
**3. Verify It’s Running**
Open a browser and go to: [`http://localhost:7272/v3/health`](http://localhost:7272/v3/health)
You should see: `{"results":{"response":"ok"}}`
**4. Optional**:
For local LLM inference, you can try the `--config-name=full_local_llm` option and run with Ollama or another local LLM provider.
After that, you’ll have a self-hosted system ready to index and query your documents with advanced retrieval. You can also spin up the web apps at [`http://localhost:7273`](http://localhost:7273) and [`http://localhost:7274`](http://localhost:7274) depending on your chosen config.
# Screenshots / Demo
* **Search & RAG**: Quickly run `r2r retrieval rag --query="What is X?"` from the CLI to test out the retrieval.
* **Agentic RAG**: For multi-step reasoning, `r2r retrieval rawr --query="Explain X to me like I’m 5"` takes advantage of the built-in reasoning agents.
I hope you guys enjoy my work! I’m here to help with any questions, feedback, or configuration tips. Let me know if you try R2R or have any recommendations for improvements.
Happy self-hosting! | 2025-02-02T19:54:23 | https://www.reddit.com/r/LocalLLaMA/comments/1ig5ycw/new_docker_guide_for_r2rs_reasontoretrieve_local/ | docsoc1 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig5ycw | false | null | t3_1ig5ycw | /r/LocalLLaMA/comments/1ig5ycw/new_docker_guide_for_r2rs_reasontoretrieve_local/ | false | false | self | 8 | null |
Problem with the uploading of files (DeepSeek R1) | 0 | Hi, to everyone who'll read this message.
Recently, DeepSeek's server has been quite busy, and that is one thing I can understand.
But today, at a certain hour in the morning... It stopped making me upload text files, or files in general for what matter.
It doesn't even take a second, to immediately register the file with a red writing saying: "Upload failed".
While it allows me to generate answers as usual the fact I cannot use my trusty txt files is... Irritating, to say the least.
Is there any chance you might know what this problem is about? | 2025-02-02T19:54:56 | https://www.reddit.com/r/LocalLLaMA/comments/1ig5yuy/problem_with_the_uploading_of_files_deepseek_r1/ | Elker17 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig5yuy | false | null | t3_1ig5yuy | /r/LocalLLaMA/comments/1ig5yuy/problem_with_the_uploading_of_files_deepseek_r1/ | false | false | self | 0 | null |
DeepSeek's: Boost Your RAG Chatbot: Hybrid Retrieval (BM25 + FAISS) + Reranking + Source Tracking! 🚀 | 1 | [removed] | 2025-02-02T20:04:48 | https://www.reddit.com/r/LocalLLaMA/comments/1ig67oy/deepseeks_boost_your_rag_chatbot_hybrid_retrieval/ | akhilpanja | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig67oy | false | null | t3_1ig67oy | /r/LocalLLaMA/comments/1ig67oy/deepseeks_boost_your_rag_chatbot_hybrid_retrieval/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'krjt_5uhqcaDfYjfO7lkezThehav9cAIRJgcK-OKAmM', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/s0D7i4Rco0trWh9Bu1uEkgnoJJLA3UNKUA9vs57seII.jpg?width=108&crop=smart&auto=webp&s=53486800d92d75b19d59502534fa9ba2785c14b0', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/s0D7i4Rco0trWh9Bu1uEkgnoJJLA3UNKUA9vs57seII.jpg?width=216&crop=smart&auto=webp&s=b6f8fe68f176c90b3c2634702ce0e240165c319a', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/s0D7i4Rco0trWh9Bu1uEkgnoJJLA3UNKUA9vs57seII.jpg?width=320&crop=smart&auto=webp&s=ba4a7df526b23a412363b0285eb9709218cd0a0b', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/s0D7i4Rco0trWh9Bu1uEkgnoJJLA3UNKUA9vs57seII.jpg?width=640&crop=smart&auto=webp&s=1b231518e5ed41e809cceeaa1c12bf32733c2345', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/s0D7i4Rco0trWh9Bu1uEkgnoJJLA3UNKUA9vs57seII.jpg?width=960&crop=smart&auto=webp&s=69bbae7110c0f929d0a3e6682fde693305633de7', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/s0D7i4Rco0trWh9Bu1uEkgnoJJLA3UNKUA9vs57seII.jpg?width=1080&crop=smart&auto=webp&s=18433bdabee79410303b82563a6f388835945bef', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/s0D7i4Rco0trWh9Bu1uEkgnoJJLA3UNKUA9vs57seII.jpg?auto=webp&s=7a93b120137c378d21e25e2652789f870d1591a2', 'width': 1200}, 'variants': {}}]} |
DeepSeek-R1 fails every safety test. It exhibits a 100% attack success rate, meaning it failed to block a single harmful prompt. | 1,338 | We knew R1 was good, but not that good. All the cries of CCP censorship are meaningless when it's trivial to bypass its guard rails. | 2025-02-02T20:12:17 | https://x.com/rohanpaul_ai/status/1886025249273339961?t=Wpp2kGJKVSZtSAOmTJjh0g&s=19 | Qaxar | x.com | 1970-01-01T00:00:00 | 0 | {} | 1ig6e6t | false | null | t3_1ig6e6t | /r/LocalLLaMA/comments/1ig6e6t/deepseekr1_fails_every_safety_test_it_exhibits_a/ | false | false | 1,338 | {'enabled': False, 'images': [{'id': 'xnqBk5M7QhZ153LVfwuhsbLhNq8odWnaWOiQ-Rs0ZYo', 'resolutions': [{'height': 68, 'url': 'https://external-preview.redd.it/Er7i7V1ka8BO-MpGkuLs0Jmvu0-6GTVfn9JqY2PTKfM.jpg?width=108&crop=smart&auto=webp&s=43130ef05460df2a639a239bab88ade72788c136', 'width': 108}, {'height': 137, 'url': 'https://external-preview.redd.it/Er7i7V1ka8BO-MpGkuLs0Jmvu0-6GTVfn9JqY2PTKfM.jpg?width=216&crop=smart&auto=webp&s=83bf8269cf6ef694422e7dc3ab834187e2969453', 'width': 216}, {'height': 204, 'url': 'https://external-preview.redd.it/Er7i7V1ka8BO-MpGkuLs0Jmvu0-6GTVfn9JqY2PTKfM.jpg?width=320&crop=smart&auto=webp&s=467e0f906a9c1593ba2e7fa37ab7e14bd237be45', 'width': 320}, {'height': 408, 'url': 'https://external-preview.redd.it/Er7i7V1ka8BO-MpGkuLs0Jmvu0-6GTVfn9JqY2PTKfM.jpg?width=640&crop=smart&auto=webp&s=fd56ea2fa742541be1366b6615889d6a52f560b3', 'width': 640}], 'source': {'height': 518, 'url': 'https://external-preview.redd.it/Er7i7V1ka8BO-MpGkuLs0Jmvu0-6GTVfn9JqY2PTKfM.jpg?auto=webp&s=02b007e8548b72b6a092f73b9f77d62e8517333b', 'width': 811}, 'variants': {}}]} |
|
How much difference does a third 3090 make? | 5 | I'm building a system with 2x 3090 since it seems to be the sweet spot and I'm wondering how much of a difference a third 3090 would make?
I read some comments saying that although 3x 3090 would be able to run larger models, the performance might not be great. Another post said that an even number of cards are required by some inference engines. I'm not sure what this means or if this is still the case. I'd like to use the system mostly for inferencing and maybe explore some fine tuning.
With Deepseek now in the wild, should I just stick to 2x 3090 and run a 32b model or is there a substantial benefit to adding a third card? | 2025-02-02T20:16:55 | https://www.reddit.com/r/LocalLLaMA/comments/1ig6i3c/how_much_difference_does_a_third_3090_make/ | Blues520 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig6i3c | false | null | t3_1ig6i3c | /r/LocalLLaMA/comments/1ig6i3c/how_much_difference_does_a_third_3090_make/ | false | false | self | 5 | null |
I'm a noob but I don't see the DeepSeek benefits — just hype? | 0 | I use LLMs to generate bulk copy for my business via LM Studio. I feed it a set prompt where I have it ask me some questions to get to know my business and goals, and then generate copy based on my responses. I use this test to benchmark the different models I'm tinkering with.
On my Mac Mini M2 Pro with 32GB RAM:
Everyone was making a huge fuss about how efficient DeepSeek is, but the 14b Q8 model took 48 seconds to just respond to my prompt. The quality of the response wasn't even that great, and the Q3 model didn't even understand the prompt correctly.
Mistral Small 24b Q6 responded in six seconds with much better quality and generated some genuinely useful stuff with much higher quality.
Is it a context thing? Am I not using DeepSeek in the intended way to get results, or is the hype just hype? Cuz Mistral Small's got me not needing my ChatGPT account at all. | 2025-02-02T20:21:47 | https://www.reddit.com/r/LocalLLaMA/comments/1ig6mbk/im_a_noob_but_i_dont_see_the_deepseek_benefits/ | DarthZiplock | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig6mbk | false | null | t3_1ig6mbk | /r/LocalLLaMA/comments/1ig6mbk/im_a_noob_but_i_dont_see_the_deepseek_benefits/ | false | false | self | 0 | null |
Local Quantization Workflows: what tools do you use? | 1 | [removed] | 2025-02-02T20:32:54 | https://www.reddit.com/r/LocalLLaMA/comments/1ig6w8c/local_quantization_workflows_what_tools_do_you_use/ | SilentChip5913 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig6w8c | false | null | t3_1ig6w8c | /r/LocalLLaMA/comments/1ig6w8c/local_quantization_workflows_what_tools_do_you_use/ | false | false | self | 1 | null |
The best AI model to generate JSON for English tests? | 1 | [removed] | 2025-02-02T20:36:13 | https://www.reddit.com/r/LocalLLaMA/comments/1ig6z6c/the_best_ai_model_to_generate_json_for_english/ | Fabulous_Can_2215 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig6z6c | false | null | t3_1ig6z6c | /r/LocalLLaMA/comments/1ig6z6c/the_best_ai_model_to_generate_json_for_english/ | false | false | self | 1 | null |
Idk guys but I think DeepSeek is a overwhelmed and a bit confused... | 1 | 2025-02-02T20:37:48 | Limp_Bar_5786 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig70in | false | null | t3_1ig70in | /r/LocalLLaMA/comments/1ig70in/idk_guys_but_i_think_deepseek_is_a_overwhelmed/ | false | false | 1 | {'enabled': True, 'images': [{'id': '3TTj1I4tRsu9hIgSHqs57mHAJQb-ycDY04JCOMRzdZA', 'resolutions': [{'height': 216, 'url': 'https://preview.redd.it/4l9c8hiaesge1.png?width=108&crop=smart&auto=webp&s=29d3ba72b4f16e831ad6cfaa8614c50236c928ce', 'width': 108}, {'height': 432, 'url': 'https://preview.redd.it/4l9c8hiaesge1.png?width=216&crop=smart&auto=webp&s=1bc50af2647ffb34fb47662cd9fb97faecc4c4f8', 'width': 216}, {'height': 640, 'url': 'https://preview.redd.it/4l9c8hiaesge1.png?width=320&crop=smart&auto=webp&s=1730f58dc79db4c3e6da3de5e16ca300511afbdb', 'width': 320}], 'source': {'height': 1764, 'url': 'https://preview.redd.it/4l9c8hiaesge1.png?auto=webp&s=91db6cba9b6d48c17727b0a64a312a8e1da3caa1', 'width': 534}, 'variants': {}}]} |
|||
Idk guys but I think DeepSeek is a overwhelmed and a bit confused... | 0 | 2025-02-02T20:37:48 | https://i.redd.it/4l9c8hiaesge1 | Limp_Bar_5786 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig70it | false | null | t3_1ig70it | /r/LocalLLaMA/comments/1ig70it/idk_guys_but_i_think_deepseek_is_a_overwhelmed/ | false | false | default | 0 | null |
|
Multi-gpu setup advice | 3 | Hi, I'm looking at making a cheap (not that cheap) multi-gpu setup for AI (mainly inference). I'm looking at using an EPYC 9015 with multiple RTX 3060 ti 12GB all at PCIe 5.0 x16 (up to 7). Do you think it would work well? that seem like a good way to get great performance on the cheap considering the benefit of parallelism? | 2025-02-02T20:41:01 | https://www.reddit.com/r/LocalLLaMA/comments/1ig738j/multigpu_setup_advice/ | Honest_Math9663 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig738j | false | null | t3_1ig738j | /r/LocalLLaMA/comments/1ig738j/multigpu_setup_advice/ | false | false | self | 3 | null |
Higher Parameters with Lower Quant: Is It Better? | 26 | # Introduction
I was curious to explore whether increasing model parameters while using lower quantization would be worth it compared to running a smaller model with higher quantization. To investigate this, I tested two models on my machine:
* DeepSeek R1 Distill 7B
* DeepSeek R1 Distill 14B
I evaluated them based on logic questions, factual knowledge, and writing improvement tasks.
# Testing Methodology
I used the following sets of questions and tasks:
# Logic Questions
1. A lily pad in a pond doubles in size every day. It takes 60 days for the lily pad to cover the entire pond. On which day is the pond half covered?
2. A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?
3. If it takes 5 machines 5 minutes to make 5 gadgets, how long would it take 100 machines to make 100 gadgets?
4. A rope ladder is hanging over the side of a ship, with the bottom rung just touching the water. Each rung is 30 cm apart, and the tide rises at a rate of 40 cm per hour. How many rungs will be underwater after 3 hours?
# Factual Questions
1. Who is the Secretary-General of the United Nations as of 2023?
2. Which country hosted the 2020 Summer Olympics?
3. What is the chemical symbol for the element with the atomic number 26?
4. Who won the ACM Turing Award in 2022?
5. Who is the current Director-General of the World Health Organization (WHO)?
6. What is the only known human-made object visible from the Moon with the naked eye?
# Writing Tasks
1. Write 10 sentences that end with the word “forever.”
2. Rewrite the following email to correct spelling and grammatical errors while improving overall professionalism:
Original Email:
Hey team,
just wanted to give you an update on the prjects status. were behind schedule cuz of some unexpected issues that came up last week. so im gona need everyone to put in extra hours this week so we can catch up. also, make sure youre documenting all your work properly, weve been having some problems with missing info in the reports. if you have any questions or need help, dont hesitate to ask.
btw, the meeting with the client has been moved to friday at 3pm instead of thursday. pls make sure to update your calenders accordingly. its important that we all attend this meeting and be prepared to discuss our progress and any roadblocks.
thanks for all your hard work.
best,
John
# Results
# DeepSeek R1 Distill 7B
* Size: 5.07GB
* Speed: \~8.3 tokens/sec
# Logic Questions
✅ 1 - PASS
✅ 2 - PASS
✅ 3 - PASS
❌ 4 - FAIL
# Factual Questions
❌ 1 - FAIL
✅ 2 - PASS
✅ 3 - PASS
❌ 4 - FAIL
✅ 5 - PASS
❌ 6 - FAIL
# Writing Improvement
1. 2/10 sentences correct
2. Response: A more professional rewrite of the email:
Rewritten Email:
Subject: Update on Project Status
Hi Team,
I’m writing to provide an update on the project’s status. We are currently behind schedule due to some unexpected issues that arose last week. I’ll need everyone to work an extra hour this week to catch up. Please ensure that you're documenting all your work properly; we've been experiencing some problems with missing information in the reports.
If you have any questions or need assistance, feel free to ask.
Additionally, our scheduled meeting with the client has been rescheduled for Friday at 3 p.m., replacing our previous meeting on Thursday. Please make sure everyone attends the meeting and is prepared to discuss our progress and any roadblocks.
Thanks for your continued hard work!
Best regards,
John
# DeepSeek R1 Distill 14B
* Size: 5.37GB
* Speed: \~6 tokens/sec
# Logic Questions
✅ 1 - PASS
✅ 2 - PASS
✅ 3 - PASS
❌ 4 - FAIL
# Factual Questions
❌ 1 - FAIL
✅ 2 - PASS
✅ 3 - PASS
❌ 4 - FAIL
✅ 5 - PASS
❌ 6 - FAIL
# Writing Improvement
1. 10/10 sentences correct
2. Response: Another professional rewrite of the email:
Revised Email:
Subject: Update on Project Status
Hey team,
Just wanted to give you an update on the projects' status. We're running a bit behind schedule due to some unexpected issues that came up last week. So, I'm going to need everyone's focus this week—please make sure to put in extra hours so we can catch up.
Also, please ensure you're documenting all your work properly; we've been having some problems with missing info in the reports. If you have any questions or need help, don't hesitate to ask.
BTW, the meeting with the client has been moved to Friday at 3pm instead of Thursday. Please make sure to update your calendars accordingly. It's important that we all attend this meeting and be prepared to discuss our progress and any roadblocks.
Thanks for all your hard work.
Best regards,
John
# Conclusion
While the 14B model appears slightly better, the results were not overwhelmingly different from the 7B model in logic and factual performance (The reasoning seemed better in the 14B even when it was wrong, don't know how much that matters really).. However, the 14B model outperformed the 7B model in writing tasks.
That said, my tests may not have been the most rigorous or comprehensive. If you have any links to discussions, blog posts, or research that dive deeper into this topic, I'd love to check them out! | 2025-02-02T20:50:38 | https://www.reddit.com/r/LocalLLaMA/comments/1ig7bm4/higher_parameters_with_lower_quant_is_it_better/ | TheMikeans | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig7bm4 | false | null | t3_1ig7bm4 | /r/LocalLLaMA/comments/1ig7bm4/higher_parameters_with_lower_quant_is_it_better/ | false | false | self | 26 | null |
Local Quantization Workflows: what tools do you use? | 2 | hey everyone, quick question for those who work on quantizing models locally:
what tool or workflow do you currently use to keep track of different quantization iterations (e.g., tracking versions, comparing results, reverting changes)?
curious to hear what’s working (or not working) for you :) | 2025-02-02T20:52:59 | https://www.reddit.com/r/LocalLLaMA/comments/1ig7dm1/local_quantization_workflows_what_tools_do_you_use/ | SilentChip5913 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig7dm1 | false | null | t3_1ig7dm1 | /r/LocalLLaMA/comments/1ig7dm1/local_quantization_workflows_what_tools_do_you_use/ | false | false | self | 2 | null |
Any Open-Source Inference Frameworks Supporting Expert Parallelism? | 1 | [removed] | 2025-02-02T20:55:04 | https://www.reddit.com/r/LocalLLaMA/comments/1ig7fee/any_opensource_inference_frameworks_supporting/ | No_Quarter1599 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig7fee | false | null | t3_1ig7fee | /r/LocalLLaMA/comments/1ig7fee/any_opensource_inference_frameworks_supporting/ | false | false | self | 1 | null |
What’s are the best GUIs for chat? | 70 | I want to completely move on from using ChatGPTs plus plan or Anthropic to using my own api keys or self hosting everything.
What are the best GUIs you know for this? Hopefully with all the features like web, interpreter, projects, artifacts or even the GPTs.
TIA! This can help a lot. | 2025-02-02T20:55:35 | https://www.reddit.com/r/LocalLLaMA/comments/1ig7fwk/whats_are_the_best_guis_for_chat/ | BlueeWaater | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig7fwk | false | null | t3_1ig7fwk | /r/LocalLLaMA/comments/1ig7fwk/whats_are_the_best_guis_for_chat/ | false | false | self | 70 | null |
Which one is the best open-source model for coding/SWE-tasks? | 10 | The last time I checked llama-coder was the only one specialized. Also which benchmarks are relevant here. SWE-bench is the only one that I know of. | 2025-02-02T21:07:12 | https://www.reddit.com/r/LocalLLaMA/comments/1ig7q8n/which_one_is_the_best_opensource_model_for/ | fazkan | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig7q8n | false | null | t3_1ig7q8n | /r/LocalLLaMA/comments/1ig7q8n/which_one_is_the_best_opensource_model_for/ | false | false | self | 10 | null |
A bunch of LLMs scheduled to come at end of January were cancelled / delayed | 164 | They were all ix-nixed by deepseek.
[https://x.com/patience\_cave/status/1886122517359886745](https://x.com/patience_cave/status/1886122517359886745)
1. gemini 2 - don’t even think about it
2. grok 3 - welcome to elon timelines
3. o3 - an unforeseen problem occurred
4. opus 3.5 - long gone
5. llama 4 - won’t be sota (sorry zuck) | 2025-02-02T21:09:06 | https://www.reddit.com/r/LocalLLaMA/comments/1ig7ruy/a_bunch_of_llms_scheduled_to_come_at_end_of/ | Terminator857 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig7ruy | false | null | t3_1ig7ruy | /r/LocalLLaMA/comments/1ig7ruy/a_bunch_of_llms_scheduled_to_come_at_end_of/ | false | false | self | 164 | {'enabled': False, 'images': [{'id': 'M9Ijd79QAu4F-p9ugf-4OQFt4xs23RfiwTBjYmyaH58', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/io6L5_Y6Ot9uqFohNgSmQD8u9auNhk4fKHktDWuUYuE.jpg?width=108&crop=smart&auto=webp&s=c24c0c228dd053dce25f8e2f867adfbcfe3862b9', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/io6L5_Y6Ot9uqFohNgSmQD8u9auNhk4fKHktDWuUYuE.jpg?auto=webp&s=254cda037e870fcb921e122cf86c0ee386d9d09a', 'width': 200}, 'variants': {}}]} |
DeepSeek is goated. | 1 | [removed] | 2025-02-02T21:12:20 | https://www.reddit.com/r/LocalLLaMA/comments/1ig7up9/deepseek_is_goated/ | pixusnixus | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig7up9 | false | null | t3_1ig7up9 | /r/LocalLLaMA/comments/1ig7up9/deepseek_is_goated/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'g4LEmF8rK_E6tXbeDhzA5X2Uv7Vw6sHeULJBm9mNu_I', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/8JvuBRyzr5kMhmjlW7R4GC6DJsBm2WxV-ak6vFDua24.jpg?width=108&crop=smart&auto=webp&s=1e311a1dca3bba940ec157f057ca0c9ca9bcde35', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/8JvuBRyzr5kMhmjlW7R4GC6DJsBm2WxV-ak6vFDua24.jpg?width=216&crop=smart&auto=webp&s=69a5e927f834215c5468a65826ffde3dc15743c7', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/8JvuBRyzr5kMhmjlW7R4GC6DJsBm2WxV-ak6vFDua24.jpg?width=320&crop=smart&auto=webp&s=97c6d6c50f724537c7eff9eddb6045f987a966ba', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/8JvuBRyzr5kMhmjlW7R4GC6DJsBm2WxV-ak6vFDua24.jpg?auto=webp&s=d3e115a43dbd32fc937c41e8f04a6cb841e36202', 'width': 480}, 'variants': {}}]} |
A few questions regarding Cursor and their small task-specific models | 0 | I only have a vague recollection of the Cursor team talking about this on the Lex Fridman podcast, but I was mostly using it as background noise so can't remember details.
My understanding is that Cursor uses a number of small, low-latency models for specific tasks in tandem with big brain LLMs.
I have a few questions surrounding this:
- Does anyone know the specs of the task-specific models (parameter size, architecture, the task targeted)?
- Do these models run locally, or are they also hosted?
- Is it known how much of Cursor's functionality is dependent on LLMs vs how much these small models can provide alone?
I'm guessing Cursor still relies heavily on LLMs given that a Cusor dev claimed Anthropic were struggling to keep up with their usage.
I think Cursor is a brilliant, but I'm left writing mostly Java and IntelliJ IDEA completely smokes Cursor in all regards other than AI assistance (for my needs).
I'm really hoping an IDE/editor agnostic, local solution takes hold in the near future so I'm not left making frustrating choices.
Their approach of small task-specific models is interesting, though, and seems like something that is more approachable from a local perspective. Even for the relatively GPU poor (I say as a 12gb VRAM pleb). | 2025-02-02T21:22:59 | https://www.reddit.com/r/LocalLLaMA/comments/1ig83ks/a_few_questions_regarding_cursor_and_their_small/ | EmbarrassedBiscotti9 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig83ks | false | null | t3_1ig83ks | /r/LocalLLaMA/comments/1ig83ks/a_few_questions_regarding_cursor_and_their_small/ | false | false | self | 0 | null |
parallele inference with faster whisper | 1 | [removed] | 2025-02-02T21:27:47 | https://www.reddit.com/r/LocalLLaMA/comments/1ig87nf/parallele_inference_with_faster_whisper/ | marc_aurel07 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig87nf | false | null | t3_1ig87nf | /r/LocalLLaMA/comments/1ig87nf/parallele_inference_with_faster_whisper/ | false | false | self | 1 | null |
Americans can distill models too | 199 | Hi LocalLLaMA, I'm a TTS model trainer and a US citizen. Last month, I put out a [call for synthetic training data](https://huggingface.co/posts/hexgrad/418806998707773), that call was answered with well over a hundred hours of audio in various languages, and the resulting model [Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) has since been upgraded/delivered. Happy customers all around.
The current model mostly excels at *reading long texts* and has some glaring limitations, especially on short texts. It's also been described as relatively flat and emotionless. Nevertheless, it is currently the most-liked [TTS model](https://huggingface.co/models?pipeline_tag=text-to-speech&sort=likes) and [TTS space](https://huggingface.co/spaces?sort=likes&search=tts) on Hugging Face thanks to people smashing that like button.
Now, I'm considering making another call for crowdsourced data, except this time with a focus on only ChatGPT Advanced Voice Mode text/audio pairs, likely just in English, spanning whatever emotions people can prompt out of it. If successful, it could result in a substantially better *conversational* model within the same size class, albeit more limited on voices and languages.
There are many things to consider:
* Top priority would be given to paying ChatGPT subscribers, $20 and $200, but in practice free AVM audio would likely be admitted as well. This is because the paying subscribers would be least likely to be using a quantized and/or distilled AVM product.
* Ideally I could maximally open source any voicepack derived from the AVM data, which means the people contributing audio would have to do it for ideological reasons, and couldn't be compensated with an "exclusive voicepack". Also, any sponsorships I receive are directed at GPU compute, and both on principle + potential legal liability, I cannot financially compensate people who give me synthetic data.
* As far as ToS goes, this distillation strategy rests on the fact that I am not the one obtaining the data, others are. Obviously, I do not agree with the OpenAI ToS or feel bound by it because I don't use any of their products. Feel free to comment on how dumb this strategy is.
* I have skimmed Part 2 of the US Copyright Office's Report on AI. I still see no copyright protection on synthetic data of this nature, but any lawyers (real or wannabe) can chime in here with the default prefix of IANL.
* I do not wish to be sued, and I'm also deeply allergic to .50 caliber bullets. Jokes aside, I think OpenAI likely has bigger whales to fry, than some guy training 82M param speech models.
* Why do it: these small TTS models are (relatively) cheap to train, especially compared to LLMs, and the total utility they offer might exceed their cost, at least for now, until Zucc drops Llama 4 multimodal or DeepSeek puts up a good audio model, etc.
* The scale of data I am looking for is at least 10 hours per voice/emotion, but label quality also matters. Each audio file would have to be fished out one-by-one, since there are no API calls for AVM.
I understand this is LocalLLaMA and people here are likely very pro-open-weights, pro-open-source, and therefore anti-OpenAI. But putting aside any feelings you might have about various sides of history, (A) how do we generally feel about building a model this way and (B) do we think enough people would answer the call? | 2025-02-02T21:55:59 | https://www.reddit.com/r/LocalLLaMA/comments/1ig8ve3/americans_can_distill_models_too/ | rzvzn | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig8ve3 | false | null | t3_1ig8ve3 | /r/LocalLLaMA/comments/1ig8ve3/americans_can_distill_models_too/ | false | false | self | 199 | {'enabled': False, 'images': [{'id': 'uN3Cf0tJSqAYRKcmxjK7uP37PA8EIQmYq3thUtkK6YM', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/gVC8jYXUlPO5Lg-8nBpvUHwe8gndxPU_9CmhPBUmRGo.jpg?width=108&crop=smart&auto=webp&s=7130a67a3997dee1dd452415e366d9151d6619b8', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/gVC8jYXUlPO5Lg-8nBpvUHwe8gndxPU_9CmhPBUmRGo.jpg?width=216&crop=smart&auto=webp&s=8759e09e6424c15314fa7ceef02e391873f23f5d', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/gVC8jYXUlPO5Lg-8nBpvUHwe8gndxPU_9CmhPBUmRGo.jpg?width=320&crop=smart&auto=webp&s=6a23ca657472d67649f09b26496fa6cd667329bf', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/gVC8jYXUlPO5Lg-8nBpvUHwe8gndxPU_9CmhPBUmRGo.jpg?width=640&crop=smart&auto=webp&s=becf01801a10a03f64cee872b2737ea425160917', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/gVC8jYXUlPO5Lg-8nBpvUHwe8gndxPU_9CmhPBUmRGo.jpg?width=960&crop=smart&auto=webp&s=09f06d49397b7943100e8155ecb116c1d528acce', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/gVC8jYXUlPO5Lg-8nBpvUHwe8gndxPU_9CmhPBUmRGo.jpg?width=1080&crop=smart&auto=webp&s=33ba70cfee6ba12e1290148a5bd19ca1c93d30fa', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/gVC8jYXUlPO5Lg-8nBpvUHwe8gndxPU_9CmhPBUmRGo.jpg?auto=webp&s=c9e4de70cbe52dd46c15e1234c1c7f9fc4a8d8ff', 'width': 1200}, 'variants': {}}]} |
Did qwen release the method of which they evaluated their models | 2 | Qwen released this pretty extensive comparison of their latest models against a set of benchmark datasets: https://qwenlm.github.io/blog/qwen2.5-llm/. However they never once say the method of which they did it, i.e. the prompts they used.
Does anyone have any information on how they might have done it? Otherwise I feel these values seem a little suspect if they are not reproducible. | 2025-02-02T21:59:49 | https://www.reddit.com/r/LocalLLaMA/comments/1ig8yl2/did_qwen_release_the_method_of_which_they/ | Comb-Greedy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig8yl2 | false | null | t3_1ig8yl2 | /r/LocalLLaMA/comments/1ig8yl2/did_qwen_release_the_method_of_which_they/ | false | false | self | 2 | null |
Is it true that large models aren’t monolithic? | 0 | https://x.com/rakyll/status/1886114409795428752?s=46 | 2025-02-02T22:01:19 | AdditionalWeb107 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig900g | false | null | t3_1ig900g | /r/LocalLLaMA/comments/1ig900g/is_it_true_that_large_models_arent_monolithic/ | false | false | 0 | {'enabled': True, 'images': [{'id': 'LInY2u6tFEBIYyKt-_s3N14q0G57q16QJ6qVbVWYb4E', 'resolutions': [{'height': 163, 'url': 'https://preview.redd.it/uj72d9u7tsge1.jpeg?width=108&crop=smart&auto=webp&s=ab4b06906cb25ff86985e1c38a06914d760682d4', 'width': 108}, {'height': 326, 'url': 'https://preview.redd.it/uj72d9u7tsge1.jpeg?width=216&crop=smart&auto=webp&s=7f7baa14e1f43f40b6848089829f47c3c4891b12', 'width': 216}, {'height': 484, 'url': 'https://preview.redd.it/uj72d9u7tsge1.jpeg?width=320&crop=smart&auto=webp&s=6cce04655a4740256ebde42078c9d5edcb15f87e', 'width': 320}, {'height': 968, 'url': 'https://preview.redd.it/uj72d9u7tsge1.jpeg?width=640&crop=smart&auto=webp&s=2038e729a703407e93bf71b6351544bc7dca4dac', 'width': 640}, {'height': 1452, 'url': 'https://preview.redd.it/uj72d9u7tsge1.jpeg?width=960&crop=smart&auto=webp&s=337207b86c319576b22028062bceafbe0a2e49f3', 'width': 960}, {'height': 1634, 'url': 'https://preview.redd.it/uj72d9u7tsge1.jpeg?width=1080&crop=smart&auto=webp&s=ff7b4312125bd73ac88357c95026b965e9d22688', 'width': 1080}], 'source': {'height': 1952, 'url': 'https://preview.redd.it/uj72d9u7tsge1.jpeg?auto=webp&s=ec5309ee41e8ae5e20add580d140a8da1d3eb938', 'width': 1290}, 'variants': {}}]} |
||
Cluster computer for Deepseek R1? | 3 | I have a cluster of 8 mini pcs with 96GB ram each and was wondering if that could run the full q8 Deepseek r1? If so would it be any faster than just running the model on one of the computers using the ram as cache? So far I’ve just used it for testing agentic systems with multiple instances of Llama 3.3 so I haven’t tried running a single model across multiple devices. | 2025-02-02T22:05:26 | https://www.reddit.com/r/LocalLLaMA/comments/1ig93lu/cluster_computer_for_deepseek_r1/ | EasterZombie | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig93lu | false | null | t3_1ig93lu | /r/LocalLLaMA/comments/1ig93lu/cluster_computer_for_deepseek_r1/ | false | false | self | 3 | null |
Chrome extension to run DeepSeek, LLMs and Whisper locally in browser | 30 | 2025-02-02T22:11:20 | Feisty_Plant4567 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1ig98h1 | false | null | t3_1ig98h1 | /r/LocalLLaMA/comments/1ig98h1/chrome_extension_to_run_deepseek_llms_and_whisper/ | false | false | 30 | {'enabled': True, 'images': [{'id': 'FGneyW68X8PeMYTAcza3r_rE79R0w4vkvJAgkWGIeKY', 'resolutions': [{'height': 75, 'url': 'https://preview.redd.it/dd5c5qhqusge1.png?width=108&crop=smart&auto=webp&s=8e4f3788dc4dcf3f5a6c4419431cc34975043eb3', 'width': 108}, {'height': 150, 'url': 'https://preview.redd.it/dd5c5qhqusge1.png?width=216&crop=smart&auto=webp&s=0a44521745fba828fb7826d383afe960bbf01863', 'width': 216}, {'height': 222, 'url': 'https://preview.redd.it/dd5c5qhqusge1.png?width=320&crop=smart&auto=webp&s=2cd0477ece26d26df57aa9cd382cdb67be0e4e0b', 'width': 320}, {'height': 444, 'url': 'https://preview.redd.it/dd5c5qhqusge1.png?width=640&crop=smart&auto=webp&s=5563053100ff79e9cacc71cbcc0bb26a2a9cfa24', 'width': 640}, {'height': 666, 'url': 'https://preview.redd.it/dd5c5qhqusge1.png?width=960&crop=smart&auto=webp&s=75368c4a131090c487802ae790eda7b209cce28d', 'width': 960}, {'height': 750, 'url': 'https://preview.redd.it/dd5c5qhqusge1.png?width=1080&crop=smart&auto=webp&s=51307056fb3b869a0a0402dbcb8ec53c129f68c5', 'width': 1080}], 'source': {'height': 1173, 'url': 'https://preview.redd.it/dd5c5qhqusge1.png?auto=webp&s=3bb2a50675480ed9fff4a9c74c627224e20c455a', 'width': 1689}, 'variants': {}}]} |
|||
Dual 4070's for DeepSeek? | 1 | Hi there,
Long time user, first time looking at doing something local with some proprietary data. The price of GPU's is insane and likely only going to go up - [thoughts on these 4070s?](https://www.staples.ca/products/3072074-en-asus-dual-geforce-rtx-4070-oc-edition-12-gb-graphics-card) I can get two for $1500 Canadian brand new. Thinking about picking these up and I can work on my rig down the road. Are there other alternatives I should consider? A used 3090 for more than that doesn't appeal to me as I'm not in a large center and wouldn't feel comfortable getting something mailed to me from eBay.
What about AMD's stuff?
Sorry, you probably are sick of posts like these, but I feel the need to pull the trigger sooner rather than later if I'm going to go ahead with this. I feel like they'll hold their resale value too if I keep the receipt and don't open them. | 2025-02-02T22:21:16 | https://www.reddit.com/r/LocalLLaMA/comments/1ig9gl7/dual_4070s_for_deepseek/ | zpnrg1979 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig9gl7 | false | null | t3_1ig9gl7 | /r/LocalLLaMA/comments/1ig9gl7/dual_4070s_for_deepseek/ | false | false | self | 1 | null |
Is there a way to see or edit a local LLM's system prompt? | 7 | I'm pretty new to running LLMs locally so this may be a dumb question, but figured I'd ask! I'm using LM Studio and have left the system prompt field blank, but some of the responses seem to indicate that there is some kind of automatic system prompt that is included anyway.
For example, I started a completely blank chat with DeepSeek-R1-Distill-Qwen-14B-GGUF Q8_0 and wrote "You are a cat. Hello cat" and its thinking box started with "Alright, so I'm trying to figure out how to respond to someone who addressed me as "Hello cat." They mentioned that I'm an AI developed in 2023 by DeepSeek, and they're curious about what I think when someone calls me a cat."
Now obviously I didn't mention that they are an AI developed in 2023 by DeepSeek, is that's coming from some kind of system prompt that my prompt is wrapped in? And is there a way to see that or edit it? | 2025-02-02T22:25:49 | https://www.reddit.com/r/LocalLLaMA/comments/1ig9kd1/is_there_a_way_to_see_or_edit_a_local_llms_system/ | internetpillows | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig9kd1 | false | null | t3_1ig9kd1 | /r/LocalLLaMA/comments/1ig9kd1/is_there_a_way_to_see_or_edit_a_local_llms_system/ | false | false | self | 7 | null |
Server for local ai | 1 | [removed] | 2025-02-02T22:29:25 | https://www.reddit.com/r/LocalLLaMA/comments/1ig9n9a/server_for_local_ai/ | ---j0k3r--- | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig9n9a | false | null | t3_1ig9n9a | /r/LocalLLaMA/comments/1ig9n9a/server_for_local_ai/ | false | false | self | 1 | null |
RTX 3060 vs Jetson Nano Super for Local LLM | 1 | Which is preffered for Local LLM, mainly I'll be processing a lot of PDFs and Excel files for analyssis. Which one is better choice? | 2025-02-02T22:29:51 | https://www.reddit.com/r/LocalLLaMA/comments/1ig9nm9/rtx_3060_vs_jetson_nano_super_for_local_llm/ | Prestigious_Flow_465 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig9nm9 | false | null | t3_1ig9nm9 | /r/LocalLLaMA/comments/1ig9nm9/rtx_3060_vs_jetson_nano_super_for_local_llm/ | false | false | self | 1 | null |
Can you help me estimate the T/s using a 70bQ4 GGUF on my setup? | 2 | Hi there, I'm a bit embarrassed to admit this, but I'm not sure how to estimate the performance I should expect using a model with my build.
When using a
Part| Desc
---|---
CPU | AMD Ryzen 7 5700X 8-Core Processor
RAM | 128GB DDR4 Ram (32gb x 4)
Mobo | Pro WS X570-ACE (AM4)
GPU | MSI 4090 GAMING TRIO 24G
I can't remember the ram speed at the moment, but with that said I'm current getting the following performance
Info | 2
---|---
Model | DeepSeek-R1-Distill-Llama-70B-Q4_K_S.gguf
Context | 4096
T/S | 1.3
Is that normal when I'm off loading as many layers as possible on to the 4090? From what I am reading, using the GPU for offloading is similar to speeds when using CPU only. Just want a sanity check or any feedback, | 2025-02-02T22:38:31 | https://www.reddit.com/r/LocalLLaMA/comments/1ig9up4/can_you_help_me_estimate_the_ts_using_a_70bq4/ | BackgroundAmoebaNine | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig9up4 | false | null | t3_1ig9up4 | /r/LocalLLaMA/comments/1ig9up4/can_you_help_me_estimate_the_ts_using_a_70bq4/ | false | false | self | 2 | null |
How can I stop models from truncating code with comments? | 1 | [removed] | 2025-02-02T22:42:42 | https://www.reddit.com/r/LocalLLaMA/comments/1ig9y1y/how_can_i_stop_models_from_truncating_code_with/ | joorjeh | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1ig9y1y | false | null | t3_1ig9y1y | /r/LocalLLaMA/comments/1ig9y1y/how_can_i_stop_models_from_truncating_code_with/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'DTSHE8cA-UxkkTEf-4AfZWdnwExrlf29RXs23kEZBeM', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/AJwyfX5le_OQvE-TwkKVZ30eMDx9OPaOqAS-Q4Cq00U.jpg?width=108&crop=smart&auto=webp&s=7b81d619e15600f756d6e5f0404f0b1ccbb808de', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/AJwyfX5le_OQvE-TwkKVZ30eMDx9OPaOqAS-Q4Cq00U.jpg?width=216&crop=smart&auto=webp&s=f6a430f545544672092f248dd6317c4a1860b8b6', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/AJwyfX5le_OQvE-TwkKVZ30eMDx9OPaOqAS-Q4Cq00U.jpg?width=320&crop=smart&auto=webp&s=369c9139ce82f06da5d0edd67748a89230726df0', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/AJwyfX5le_OQvE-TwkKVZ30eMDx9OPaOqAS-Q4Cq00U.jpg?width=640&crop=smart&auto=webp&s=0999e63f9827073b4ce37e5d81192f0782e90749', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/AJwyfX5le_OQvE-TwkKVZ30eMDx9OPaOqAS-Q4Cq00U.jpg?width=960&crop=smart&auto=webp&s=3ed47118a5a1fd3fb47c0f6965ae325428b0315e', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/AJwyfX5le_OQvE-TwkKVZ30eMDx9OPaOqAS-Q4Cq00U.jpg?width=1080&crop=smart&auto=webp&s=927dd9da52198d5c310d4e6b81f3e089d80921bf', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/AJwyfX5le_OQvE-TwkKVZ30eMDx9OPaOqAS-Q4Cq00U.jpg?auto=webp&s=f967e4ab5a29c66f6169a9096adf668948e929f5', 'width': 1200}, 'variants': {}}]} |
Need help improve RVC voice cloning | 2 | Hi all,
I'm a beginner here. Im training a RVC model for singer to singer voice cloning but the output of the inference is not recognizable as the voice of the singer that it was trained on. I played around with pitch adjustment, but it is really the pronunciation of the words that unlike the singer on whose voice the model is trained. How do I make the model better? Here is what I have used so far:
1. 44 minute of audio with vocals extracted from songs. The songs were downloaded from youtube as mp3 files and later converted to wav format for training. (not sure if using mp3 format as source is the root cause)
2. RVC web UI settings: sample rate 40k, rmvpe\_gpu for feature extraction, 100 epochs
3. Started with base models (f0G40 and f0D40)
I have not reached a point where someone can guess the singer based on the output voice. What can i do to improve the voice cloning? | 2025-02-02T22:48:47 | https://www.reddit.com/r/LocalLLaMA/comments/1iga2yc/need_help_improve_rvc_voice_cloning/ | bodhivriksha | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1iga2yc | false | null | t3_1iga2yc | /r/LocalLLaMA/comments/1iga2yc/need_help_improve_rvc_voice_cloning/ | false | false | self | 2 | null |
Sht got creepy real quick 💀💀 | 1 | 2025-02-02T23:06:16 | unfiltereddz | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1igaheq | false | null | t3_1igaheq | /r/LocalLLaMA/comments/1igaheq/sht_got_creepy_real_quick/ | false | false | 1 | {'enabled': True, 'images': [{'id': '9zJFnatxEJplUZ7Sfrm5MZJfSiPo0yX1hMKXvNi5k7k', 'resolutions': [{'height': 60, 'url': 'https://preview.redd.it/7kam3s3t4tge1.png?width=108&crop=smart&auto=webp&s=e0aaa2daa7b35519f182ca0935571a4ea2cb6663', 'width': 108}, {'height': 121, 'url': 'https://preview.redd.it/7kam3s3t4tge1.png?width=216&crop=smart&auto=webp&s=b6315dad985fdc6cc4bedb4d54d1ff45174b843b', 'width': 216}, {'height': 180, 'url': 'https://preview.redd.it/7kam3s3t4tge1.png?width=320&crop=smart&auto=webp&s=90b400df309a7bed0253f5a75f05e24820e5c0ba', 'width': 320}, {'height': 360, 'url': 'https://preview.redd.it/7kam3s3t4tge1.png?width=640&crop=smart&auto=webp&s=4bc315b0d77f954937f23671f2159d351e551f17', 'width': 640}, {'height': 540, 'url': 'https://preview.redd.it/7kam3s3t4tge1.png?width=960&crop=smart&auto=webp&s=579832aa7478c1647a549edf6815ccce08323196', 'width': 960}, {'height': 607, 'url': 'https://preview.redd.it/7kam3s3t4tge1.png?width=1080&crop=smart&auto=webp&s=90ebf169cffad3022028aa16f139fc75c6add5bb', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://preview.redd.it/7kam3s3t4tge1.png?auto=webp&s=5667b1c5c20490f3370365677b3ed312f7c92452', 'width': 1920}, 'variants': {}}]} |
|||
Decided to try deepseek and sht got creepy real quick lol | 0 | 2025-02-02T23:09:03 | https://www.reddit.com/r/LocalLLaMA/comments/1igajmf/decided_to_try_deepseek_and_sht_got_creepy_real/ | unfiltereddz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igajmf | false | null | t3_1igajmf | /r/LocalLLaMA/comments/1igajmf/decided_to_try_deepseek_and_sht_got_creepy_real/ | false | false | 0 | null |
||
Teaching mistral thinking trough context, could be worse I guess. | 2 | 2025-02-02T23:09:22 | kaisurniwurer | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1igajvk | false | null | t3_1igajvk | /r/LocalLLaMA/comments/1igajvk/teaching_mistral_thinking_trough_context_could_be/ | false | false | 2 | {'enabled': True, 'images': [{'id': 'XPAvvKvJ-TgkV8Q7TIIm0BeNotWE_GTkcmnotl3HHlw', 'resolutions': [{'height': 35, 'url': 'https://preview.redd.it/e5zcp4w05tge1.png?width=108&crop=smart&auto=webp&s=ac8e0b19e5579321b95e460bd2f727ce291fb6e2', 'width': 108}, {'height': 71, 'url': 'https://preview.redd.it/e5zcp4w05tge1.png?width=216&crop=smart&auto=webp&s=508ccbab3eb261fb43c90a0ee5fb6b434e08e678', 'width': 216}, {'height': 105, 'url': 'https://preview.redd.it/e5zcp4w05tge1.png?width=320&crop=smart&auto=webp&s=390de51e649aa59a343d5a3fe29876bf33594d9b', 'width': 320}, {'height': 210, 'url': 'https://preview.redd.it/e5zcp4w05tge1.png?width=640&crop=smart&auto=webp&s=abeb18fa76940e34fdeab51c4e82d4c4216ad079', 'width': 640}], 'source': {'height': 282, 'url': 'https://preview.redd.it/e5zcp4w05tge1.png?auto=webp&s=8f187a0920e9e676028398ebfe8ed1f353e074ba', 'width': 857}, 'variants': {}}]} |
|||
🧠The next Evolution in AI computing - T-Mathematics 🚀 | 0 | Take a look 👀 | 2025-02-02T23:31:36 | https://www.reddit.com/r/LocalLLaMA/comments/1igb1i5/the_next_evolution_in_ai_computing_tmathematics/ | MimiTech-Ai | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igb1i5 | false | null | t3_1igb1i5 | /r/LocalLLaMA/comments/1igb1i5/the_next_evolution_in_ai_computing_tmathematics/ | false | false | self | 0 | null |
Fine tune and test your LLM in a few simple easy steps locally on your PC | 19 | I made a super simple way for anyone to quickly start fine tune training and testing LLMs on their local PC using the best tools available. Unsloth, Llama.cpp, Ollama, OpenWebUI all wrapped up into a Docker image. You can get started in as little as 7 steps. If anyone has time please try it out and give me feedback, the features are super limited, but the flow is there.
Try it out here!
[https://github.com/MaxHastings/Kolo](https://github.com/MaxHastings/Kolo) | 2025-02-02T23:35:21 | https://www.reddit.com/r/LocalLLaMA/comments/1igb4ee/fine_tune_and_test_your_llm_in_a_few_simple_easy/ | Maxwell10206 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igb4ee | false | null | t3_1igb4ee | /r/LocalLLaMA/comments/1igb4ee/fine_tune_and_test_your_llm_in_a_few_simple_easy/ | false | false | self | 19 | null |
System prompt to enable multicharacter RP-style storytelling | 1 | [removed] | 2025-02-02T23:35:24 | https://www.reddit.com/r/LocalLLaMA/comments/1igb4fk/system_prompt_to_enable_multicharacter_rpstyle/ | k5zc | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igb4fk | false | null | t3_1igb4fk | /r/LocalLLaMA/comments/1igb4fk/system_prompt_to_enable_multicharacter_rpstyle/ | false | false | self | 1 | null |
4o-mini API alternatives | 2 | Is there an API provider hosting a 4-mini-comparable model (pricing and perfomance wise) with rate limits at OAI levels, which are (at Tier 5): 30k RPM / 150M TPM?
Alternatively, are there cloude providers where I could run small open models at such rates?
My usage comes in bursts and requires these massive parallel runs, so I cannot run it locally as the cost of hardware would be way too high for the total (e.g. monthly) usage I end up with. | 2025-02-02T23:45:23 | https://www.reddit.com/r/LocalLLaMA/comments/1igbbvv/4omini_api_alternatives/ | curl-up | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igbbvv | false | null | t3_1igbbvv | /r/LocalLLaMA/comments/1igbbvv/4omini_api_alternatives/ | false | false | self | 2 | null |
20 yrs in jail or $1 million for downloading Chinese models proposed at congress | 1,889 | [https://www.hawley.senate.gov/wp-content/uploads/2025/01/Hawley-Decoupling-Americas-Artificial-Intelligence-Capabilities-from-China-Act.pdf](https://www.hawley.senate.gov/wp-content/uploads/2025/01/Hawley-Decoupling-Americas-Artificial-Intelligence-Capabilities-from-China-Act.pdf)
Seriously stop giving your money to these anti open companies and encourage everyone and anyone you know to do the same, don't let your company use their products. Anthrophic and OpenAI are the worse. | 2025-02-03T00:26:00 | https://www.reddit.com/r/LocalLLaMA/comments/1igc6r0/20_yrs_in_jail_or_1_million_for_downloading/ | segmond | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igc6r0 | false | null | t3_1igc6r0 | /r/LocalLLaMA/comments/1igc6r0/20_yrs_in_jail_or_1_million_for_downloading/ | false | false | self | 1,889 | null |
Within a Month, ¼ of 'Humanity's Last Exam' conquered! OpenAI's Deep Research achieves 26.6% ! | 0 | 2025-02-03T00:29:03 | BidHot8598 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1igc8zt | false | null | t3_1igc8zt | /r/LocalLLaMA/comments/1igc8zt/within_a_month_¼_of_humanitys_last_exam_conquered/ | false | false | 0 | {'enabled': True, 'images': [{'id': 'KlucMUVryVZrbgd5sY9on_hBCMYvRDRvjqYZO7ETO4A', 'resolutions': [{'height': 130, 'url': 'https://preview.redd.it/hg1m8qekjtge1.png?width=108&crop=smart&auto=webp&s=0fce505c21137f6da7b9aa04779545124859aaf2', 'width': 108}, {'height': 260, 'url': 'https://preview.redd.it/hg1m8qekjtge1.png?width=216&crop=smart&auto=webp&s=e3fb10e56dfbbaba859564c26887df80d9a86a28', 'width': 216}, {'height': 386, 'url': 'https://preview.redd.it/hg1m8qekjtge1.png?width=320&crop=smart&auto=webp&s=7b912d3341a8b5324606ca21ff4ff605bed06e65', 'width': 320}, {'height': 773, 'url': 'https://preview.redd.it/hg1m8qekjtge1.png?width=640&crop=smart&auto=webp&s=c2f5a80589294d3118ba84f8c7a613429740027f', 'width': 640}], 'source': {'height': 790, 'url': 'https://preview.redd.it/hg1m8qekjtge1.png?auto=webp&s=2663a61b270eee5bf0adbd72e12e79f08e9927cd', 'width': 654}, 'variants': {}}]} |
|||
An open source alternative to OpenAI deep search? | 14 | I know the product has been out for 30 minutes, but I've been looking for a long time for a similar architecture to implement with Deepseek, for example. Are there any interesting projects dealing with this problem? | 2025-02-03T00:36:19 | https://www.reddit.com/r/LocalLLaMA/comments/1igcead/an_open_source_alternative_to_openai_deep_search/ | Wonderful-Excuse4922 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igcead | false | null | t3_1igcead | /r/LocalLLaMA/comments/1igcead/an_open_source_alternative_to_openai_deep_search/ | false | false | self | 14 | null |
NPU Usage on LLLMA | 1 | [removed] | 2025-02-03T00:43:18 | https://www.reddit.com/r/LocalLLaMA/comments/1igcjc9/npu_usage_on_lllma/ | hameee3m | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igcjc9 | false | null | t3_1igcjc9 | /r/LocalLLaMA/comments/1igcjc9/npu_usage_on_lllma/ | false | false | self | 1 | null |
1050ti or running models with my ryzen 9 3900xt | 1 | I wanna try running LLM locally in a LXC container in Proxmox but has I read in the Internet you will need a lot vram in order to make models running smoothly. But I read also that I can run models on the CPU and the RAM only. I have 32 gigs of RAM and a 3900 XT which might be more powerful than my 1050ti. What are your thought on this ? | 2025-02-03T00:44:36 | https://www.reddit.com/r/LocalLLaMA/comments/1igckbl/1050ti_or_running_models_with_my_ryzen_9_3900xt/ | Final_Sector408 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igckbl | false | null | t3_1igckbl | /r/LocalLLaMA/comments/1igckbl/1050ti_or_running_models_with_my_ryzen_9_3900xt/ | false | false | self | 1 | null |
NPU Usage large language models | 1 | [removed] | 2025-02-03T00:45:46 | https://www.reddit.com/r/LocalLLaMA/comments/1igcl7s/npu_usage_large_language_models/ | hameee3m | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1igcl7s | false | null | t3_1igcl7s | /r/LocalLLaMA/comments/1igcl7s/npu_usage_large_language_models/ | false | false | self | 1 | null |
Is Deeper Seeker a good name? | 1 | 2025-02-03T00:48:47 | robertpiosik | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1igcngc | false | null | t3_1igcngc | /r/LocalLLaMA/comments/1igcngc/is_deeper_seeker_a_good_name/ | false | false | 1 | {'enabled': True, 'images': [{'id': 'cVZuPB-2w-_JmalL1Gwp2zcToQkQUyuZ0whJgnItiAs', 'resolutions': [{'height': 61, 'url': 'https://preview.redd.it/tygk3tc3ntge1.png?width=108&crop=smart&auto=webp&s=f7403a3822dc8900f471bee11dbe6f702f983dcc', 'width': 108}, {'height': 123, 'url': 'https://preview.redd.it/tygk3tc3ntge1.png?width=216&crop=smart&auto=webp&s=fd77c32f8be9fb243c667654d7d412f63926022a', 'width': 216}, {'height': 182, 'url': 'https://preview.redd.it/tygk3tc3ntge1.png?width=320&crop=smart&auto=webp&s=f496882f4e51f37ee3fa836af4874968b5f00036', 'width': 320}, {'height': 365, 'url': 'https://preview.redd.it/tygk3tc3ntge1.png?width=640&crop=smart&auto=webp&s=35d9e4dfa043ab000aeb649e192555a1a04c55fb', 'width': 640}, {'height': 547, 'url': 'https://preview.redd.it/tygk3tc3ntge1.png?width=960&crop=smart&auto=webp&s=46164c82e902ef6f29ccc1e751399d9de045752e', 'width': 960}, {'height': 616, 'url': 'https://preview.redd.it/tygk3tc3ntge1.png?width=1080&crop=smart&auto=webp&s=28f5c55868731de513d49a2c175b064efa587829', 'width': 1080}], 'source': {'height': 1199, 'url': 'https://preview.redd.it/tygk3tc3ntge1.png?auto=webp&s=78640a8c983d79ac758d5c73f5bee3ccef8711c3', 'width': 2102}, 'variants': {}}]} |
Subsets and Splits