title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns]date 2023-04-01 04:30:41
2025-06-30 03:16:29
⌀ | url
stringlengths 0
878
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns]date 1970-01-01 00:00:00
2025-06-26 17:30:18
| gilded
int64 0
2
| gildings
stringclasses 7
values | id
stringlengths 7
7
| locked
bool 2
classes | media
stringlengths 646
1.8k
⌀ | name
stringlengths 10
10
| permalink
stringlengths 33
82
| spoiler
bool 2
classes | stickied
bool 2
classes | thumbnail
stringlengths 4
213
| ups
int64 0
8.54k
| preview
stringlengths 301
5.01k
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Unsloth accuracy vs Transformers | 1 | [removed] | 2025-01-22T02:11:41 | https://www.reddit.com/r/LocalLLaMA/comments/1i7056h/unsloth_accuracy_vs_transformers/ | BitAcademic9597 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7056h | false | null | t3_1i7056h | /r/LocalLLaMA/comments/1i7056h/unsloth_accuracy_vs_transformers/ | false | false | self | 1 | null |
(any) Routing - unified access and observability for API-based and Ollama-supported LLMs in v0.1.9 | 1 | [removed] | 2025-01-22T02:23:32 | AdditionalWeb107 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i70dvu | false | null | t3_1i70dvu | /r/LocalLLaMA/comments/1i70dvu/any_routing_unified_access_and_observability_for/ | false | false | 1 | {'enabled': True, 'images': [{'id': 'ih4j-P8Gi7qMNX-zcAmCErFtVS2Tpg9Vqf6cGgWikuw', 'resolutions': [{'height': 108, 'url': 'https://preview.redd.it/17xgbf92hgee1.jpeg?width=108&crop=smart&auto=webp&s=b666c95b7a553847373b4dfbf3bef845b8dd4dc7', 'width': 108}, {'height': 216, 'url': 'https://preview.redd.it/17xgbf92hgee1.jpeg?width=216&crop=smart&auto=webp&s=456d960255c3755749106ba94528c5e29f6cc87c', 'width': 216}, {'height': 320, 'url': 'https://preview.redd.it/17xgbf92hgee1.jpeg?width=320&crop=smart&auto=webp&s=ef1016e3cf218fca2ca183041b9d60ee8daf26f8', 'width': 320}, {'height': 640, 'url': 'https://preview.redd.it/17xgbf92hgee1.jpeg?width=640&crop=smart&auto=webp&s=186b3b2e18dd48a6d941565426ab7b8bf4ea4992', 'width': 640}, {'height': 960, 'url': 'https://preview.redd.it/17xgbf92hgee1.jpeg?width=960&crop=smart&auto=webp&s=eefc174e65dc7f95a189819097b4eb8deeb93b1d', 'width': 960}, {'height': 1080, 'url': 'https://preview.redd.it/17xgbf92hgee1.jpeg?width=1080&crop=smart&auto=webp&s=8765cf35f02475ab7f86bb511550af3c0fb8c7c5', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://preview.redd.it/17xgbf92hgee1.jpeg?auto=webp&s=20feda324f28b3d711812939164466433a7172d4', 'width': 1080}, 'variants': {}}]} |
||
Any Router - unified access and observability for API-based and Ollama-supported LLMs | 1 | [removed] | 2025-01-22T02:33:25 | AdditionalWeb107 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i70kyv | false | null | t3_1i70kyv | /r/LocalLLaMA/comments/1i70kyv/any_router_unified_access_and_observability_for/ | false | false | 1 | {'enabled': True, 'images': [{'id': 'Z_20aEmrbtchcCMi9Pv0iCv32QoDB2ABjum8I7Igceg', 'resolutions': [{'height': 108, 'url': 'https://preview.redd.it/m78u81rtigee1.jpeg?width=108&crop=smart&auto=webp&s=c6edef64cd745998377587d1e59b046fdff81e3a', 'width': 108}, {'height': 216, 'url': 'https://preview.redd.it/m78u81rtigee1.jpeg?width=216&crop=smart&auto=webp&s=704f817f83f24483669314bf98e0ba0e5eaf8680', 'width': 216}, {'height': 320, 'url': 'https://preview.redd.it/m78u81rtigee1.jpeg?width=320&crop=smart&auto=webp&s=a295d826273effffdd91d36a08a56226f9868246', 'width': 320}, {'height': 640, 'url': 'https://preview.redd.it/m78u81rtigee1.jpeg?width=640&crop=smart&auto=webp&s=713d3652032092ae61a995483a86d06548ffec2b', 'width': 640}, {'height': 960, 'url': 'https://preview.redd.it/m78u81rtigee1.jpeg?width=960&crop=smart&auto=webp&s=59d93d4694d5ff28051a7562755e61c64d42d001', 'width': 960}, {'height': 1080, 'url': 'https://preview.redd.it/m78u81rtigee1.jpeg?width=1080&crop=smart&auto=webp&s=e973abb789396f1942e20cdad43999c2db945d84', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://preview.redd.it/m78u81rtigee1.jpeg?auto=webp&s=f023f7ee24fb41230a8ebd40e864cf84858f8a73', 'width': 1080}, 'variants': {}}]} |
||
Are there scenarios where your application need access to local LLMs and a API-based ones? | 3 | The title says it all. Would be curious to get your thoughts on the usage of local LLMs and API-based ones through your application. I am seeing a lot of folks experiment with different models as they emerge and were curious if the community is mixing/matching models for specific application use cases. | 2025-01-22T02:36:20 | https://www.reddit.com/r/LocalLLaMA/comments/1i70n2t/are_there_scenarios_where_your_application_need/ | AdditionalWeb107 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i70n2t | false | null | t3_1i70n2t | /r/LocalLLaMA/comments/1i70n2t/are_there_scenarios_where_your_application_need/ | false | false | self | 3 | null |
Unsloth is fast and easy but is it accurate? (for professional tasks) | 1 | [removed] | 2025-01-22T02:37:12 | https://www.reddit.com/r/LocalLLaMA/comments/1i70noo/unsloth_is_fast_and_easy_but_is_it_accurate_for/ | BitAcademic9597 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i70noo | false | null | t3_1i70noo | /r/LocalLLaMA/comments/1i70noo/unsloth_is_fast_and_easy_but_is_it_accurate_for/ | false | false | self | 1 | null |
R1 randomly started talking in chinese | 0 | 2025-01-22T02:43:18 | https://www.reddit.com/r/LocalLLaMA/comments/1i70sml/r1_randomly_started_talking_in_chinese/ | Reader3123 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i70sml | false | null | t3_1i70sml | /r/LocalLLaMA/comments/1i70sml/r1_randomly_started_talking_in_chinese/ | false | false | 0 | null |
||
Unsloth accuracy vs Hugging face Transformers for professional | 1 | [removed] | 2025-01-22T02:56:08 | https://www.reddit.com/r/LocalLLaMA/comments/1i712si/unsloth_accuracy_vs_hugging_face_transformers_for/ | BitAcademic9597 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i712si | false | null | t3_1i712si | /r/LocalLLaMA/comments/1i712si/unsloth_accuracy_vs_hugging_face_transformers_for/ | false | false | self | 1 | null |
Appearancly Deepseek Roleplaying Has More Freedom | 2 | [removed] | 2025-01-22T03:11:27 | https://www.reddit.com/gallery/1i71fcl | _HornyXytrek_ | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1i71fcl | false | null | t3_1i71fcl | /r/LocalLLaMA/comments/1i71fcl/appearancly_deepseek_roleplaying_has_more_freedom/ | false | false | nsfw | 2 | null |
Difference between DeepSeek and OpenAI? | 20 | How come OpenAI o1 + o1-mini has usage limits ($20 plus plan for 50 messages a week or $200 pro plan for unlimited) while DeepSeek's "DeepThink" is completely free to use on their site?
Am I missing something? Does DeepSeek's reasoning architecture have way less compute costs or does DeepSeek's reasoning not even close to what o1 can do? | 2025-01-22T03:16:29 | https://www.reddit.com/r/LocalLLaMA/comments/1i71j8q/difference_between_deepseek_and_openai/ | devinak | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i71j8q | false | null | t3_1i71j8q | /r/LocalLLaMA/comments/1i71j8q/difference_between_deepseek_and_openai/ | false | false | self | 20 | null |
Is 32x32GB RTX5090s enough for r1 and llama 70b | 0 | So they support 8 KV heads right? Would that be clusters of 4 then? I’m thinking of buying 32 rtx 5090s for a company agent system. | 2025-01-22T03:26:08 | https://www.reddit.com/r/LocalLLaMA/comments/1i71qpw/is_32x32gb_rtx5090s_enough_for_r1_and_llama_70b/ | BarnardWellesley | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i71qpw | false | null | t3_1i71qpw | /r/LocalLLaMA/comments/1i71qpw/is_32x32gb_rtx5090s_enough_for_r1_and_llama_70b/ | false | false | self | 0 | null |
“Any Router” in v0.1.9 - unify access and observability to ollama-supported and API-based LLMs | 24 | Not a new project; just a feature update on egress functionally to help developers unify access and observability for ollama- supported and API-based LLMs. Coincidentally out #1 feature request last month.
So if you want a simple way to access any LLM and get unified tracing and logs, then this update might be useful for you.
https://github.com/katanemo/archgw.
Arch Gateway is an intelligent proxy server designed for prompts. Guides for egress LLM routing and ollama below
https://github.com/katanemo/archgw?tab=readme-ov-file#use-arch-gateway-as-llm-router
https://github.com/katanemo/archgw/tree/main/demos/currency_exchange_ollamau | 2025-01-22T03:53:08 | AdditionalWeb107 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i72bip | false | null | t3_1i72bip | /r/LocalLLaMA/comments/1i72bip/any_router_in_v019_unify_access_and_observability/ | false | false | 24 | {'enabled': True, 'images': [{'id': '3otV4kSUjq1KIcebneW56_Cvr-w7a7v4lnXutC7MgKo', 'resolutions': [{'height': 108, 'url': 'https://preview.redd.it/ks10onr1xgee1.jpeg?width=108&crop=smart&auto=webp&s=56e27a9b5b05639962cfc554d9fb11afe95a4c08', 'width': 108}, {'height': 216, 'url': 'https://preview.redd.it/ks10onr1xgee1.jpeg?width=216&crop=smart&auto=webp&s=90b78b81009bbdd104779a1d383a857d8a5f139a', 'width': 216}, {'height': 320, 'url': 'https://preview.redd.it/ks10onr1xgee1.jpeg?width=320&crop=smart&auto=webp&s=8e975894f428b961a66ba94cb475ea1aa3724997', 'width': 320}, {'height': 640, 'url': 'https://preview.redd.it/ks10onr1xgee1.jpeg?width=640&crop=smart&auto=webp&s=61fada01dff806ec53c0bd3dc33cb257f94507e7', 'width': 640}, {'height': 960, 'url': 'https://preview.redd.it/ks10onr1xgee1.jpeg?width=960&crop=smart&auto=webp&s=3aa68ce31e8eba50621c476e2c8f907b612b6d21', 'width': 960}, {'height': 1080, 'url': 'https://preview.redd.it/ks10onr1xgee1.jpeg?width=1080&crop=smart&auto=webp&s=f4a6a5703a5b19f20b24ff44bf1a39f2f09b93b2', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://preview.redd.it/ks10onr1xgee1.jpeg?auto=webp&s=50a316721e2802933ed7ec7500cfbf39f4ed034f', 'width': 1080}, 'variants': {}}]} |
||
DeepSeek R1-Zero is hilariously uncensored, and now available on Hyperbolic AI. | 1 | [removed] | 2025-01-22T04:19:07 | https://www.reddit.com/r/LocalLLaMA/comments/1i72tft/deepseek_r1zero_is_hilariously_uncensored_and_now/ | georgejrjrjr | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i72tft | false | null | t3_1i72tft | /r/LocalLLaMA/comments/1i72tft/deepseek_r1zero_is_hilariously_uncensored_and_now/ | false | false | self | 1 | null |
M4 Mini Pro for Training LLMs | 2 | I recently bought an M4 Mini, as a replacement for my old laptop, to run and train LLMs, locally.
I just wanted to know if my current specs would be enough, and what few configurations people would recommend for this.
Specs:
24GB Unified Memory
512GB SSD
12-core CPU 16-core GPU | 2025-01-22T04:23:10 | https://www.reddit.com/r/LocalLLaMA/comments/1i72w5b/m4_mini_pro_for_training_llms/ | Scapegoat079 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i72w5b | false | null | t3_1i72w5b | /r/LocalLLaMA/comments/1i72w5b/m4_mini_pro_for_training_llms/ | false | false | self | 2 | null |
Missing vision capability or did we just uncover the truth? | 0 | 2025-01-22T04:24:42 | https://www.reddit.com/r/LocalLLaMA/comments/1i72x4o/missing_vision_capability_or_did_we_just_uncover/ | thedgyalt | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i72x4o | false | null | t3_1i72x4o | /r/LocalLLaMA/comments/1i72x4o/missing_vision_capability_or_did_we_just_uncover/ | false | false | 0 | null |
||
3090 nvlink… is there a point? | 3 | I’ve got two 3090s (one in a gaming pc and the other sitting around until I can get it to a family member)… I’m wondering is there a point to pooling the vram with nvlink to have better models running vs even a 32gb 5090?
Curious on people’s thoughts | 2025-01-22T04:37:17 | https://www.reddit.com/r/LocalLLaMA/comments/1i7359q/3090_nvlink_is_there_a_point/ | the_shek | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7359q | false | null | t3_1i7359q | /r/LocalLLaMA/comments/1i7359q/3090_nvlink_is_there_a_point/ | false | false | self | 3 | null |
What are the commercial benefits from the opensource deepseek reasoning models ? | 5 | With all this better benchmark which we are looking from deepseek. As it is opensource , my question is apart from research purpose where exactly you guys can use those reasoning opensource models as there is thinking/latency in every response. I would love to know the best use cases of opensource reasoning models. | 2025-01-22T04:46:54 | https://www.reddit.com/r/LocalLLaMA/comments/1i73b5q/what_are_the_commercial_benefits_from_the/ | Maleficent_Repair359 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i73b5q | false | null | t3_1i73b5q | /r/LocalLLaMA/comments/1i73b5q/what_are_the_commercial_benefits_from_the/ | false | false | self | 5 | null |
Would it be better to download a highly quantized large model, or a raw small model? | 2 | Title, basically. Does the large model have enough base quality that the quantization beats the raw smaller version? | 2025-01-22T04:50:45 | https://www.reddit.com/r/LocalLLaMA/comments/1i73dhd/would_it_be_better_to_download_a_highly_quantized/ | Donovanth1 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i73dhd | false | null | t3_1i73dhd | /r/LocalLLaMA/comments/1i73dhd/would_it_be_better_to_download_a_highly_quantized/ | false | false | self | 2 | null |
AskDS - automatically send test failures and your repo to R1 to get actionable results to fix the failing tests | 1 | 2025-01-22T04:57:54 | https://github.com/mohsen1/askds | mehyay76 | github.com | 1970-01-01T00:00:00 | 0 | {} | 1i73huu | false | null | t3_1i73huu | /r/LocalLLaMA/comments/1i73huu/askds_automatically_send_test_failures_and_your/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'CTi3I_IW8nH6aJxzbqlgAvtM6y0MNN-Ood3IpxOn9c8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/4pRQcaVlgTffkpbnHAL3O7iqms1Xvjb1fXS6BGA6gq4.jpg?width=108&crop=smart&auto=webp&s=56329002a5450d890c093ed90760fe83e9bcaa58', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/4pRQcaVlgTffkpbnHAL3O7iqms1Xvjb1fXS6BGA6gq4.jpg?width=216&crop=smart&auto=webp&s=809725a2ff2057907daf852629541d6795a6e249', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/4pRQcaVlgTffkpbnHAL3O7iqms1Xvjb1fXS6BGA6gq4.jpg?width=320&crop=smart&auto=webp&s=ab85244ee280dfe7351fb11a7f8dd0f902fda8e3', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/4pRQcaVlgTffkpbnHAL3O7iqms1Xvjb1fXS6BGA6gq4.jpg?width=640&crop=smart&auto=webp&s=73261ac199f66685228b3308c4ffc106f07bcd68', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/4pRQcaVlgTffkpbnHAL3O7iqms1Xvjb1fXS6BGA6gq4.jpg?width=960&crop=smart&auto=webp&s=2b26f89bc3c76ab79090efa6f078d8a38cd025a2', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/4pRQcaVlgTffkpbnHAL3O7iqms1Xvjb1fXS6BGA6gq4.jpg?width=1080&crop=smart&auto=webp&s=8cfa5a88d89d7b07c15b534a4f8fe60c0aa3a430', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/4pRQcaVlgTffkpbnHAL3O7iqms1Xvjb1fXS6BGA6gq4.jpg?auto=webp&s=aef52d5e79b2adc5e189d584e49adaf887722456', 'width': 1200}, 'variants': {}}]} |
||
AskDS - automatically send test failures and your repo to R1 to get actionable results to fix the failing test | 10 | 2025-01-22T05:00:16 | https://v.redd.it/l0zurgqy8hee1 | mehyay76 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i73jcv | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/l0zurgqy8hee1/DASHPlaylist.mpd?a=1740114032%2CZjYxOGUwZGUyN2VkY2E0YWVkYzJmYTA5NDI2MjYxNTgzYjdlZGVlZjM0NjQ0YTMwNGFiY2Q0MmRhYmZmMTdkOQ%3D%3D&v=1&f=sd', 'duration': 45, 'fallback_url': 'https://v.redd.it/l0zurgqy8hee1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/l0zurgqy8hee1/HLSPlaylist.m3u8?a=1740114032%2CODMxM2QyZTcxN2YyODBiNGY2N2I1MTAyNzhmY2FkZjYwMmEyNTJjNDUxMDk3ZGIyNDMxMjI2NTkxNzgwZGRlMQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/l0zurgqy8hee1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1428}} | t3_1i73jcv | /r/LocalLLaMA/comments/1i73jcv/askds_automatically_send_test_failures_and_your/ | false | false | 10 | {'enabled': False, 'images': [{'id': 'dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP.png?width=108&crop=smart&format=pjpg&auto=webp&s=d8a63e70c0dbd79c57aee0abceda24fcd7f78a37', 'width': 108}, {'height': 163, 'url': 'https://external-preview.redd.it/dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP.png?width=216&crop=smart&format=pjpg&auto=webp&s=3f0ab9e1b604fecb1ee76cd405a8679bb168c731', 'width': 216}, {'height': 242, 'url': 'https://external-preview.redd.it/dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP.png?width=320&crop=smart&format=pjpg&auto=webp&s=727274a459ad810aa1b910df2263c864b319a2cb', 'width': 320}, {'height': 484, 'url': 'https://external-preview.redd.it/dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP.png?width=640&crop=smart&format=pjpg&auto=webp&s=934efc8235c22ee19345ec6f3f7c023367c9b9d8', 'width': 640}, {'height': 726, 'url': 'https://external-preview.redd.it/dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP.png?width=960&crop=smart&format=pjpg&auto=webp&s=cacea50e74f09a43131f85ecd1c29a0d8455f923', 'width': 960}, {'height': 817, 'url': 'https://external-preview.redd.it/dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP.png?width=1080&crop=smart&format=pjpg&auto=webp&s=233270331436a5b9d3c1629c9b1f89157eef8a96', 'width': 1080}], 'source': {'height': 1342, 'url': 'https://external-preview.redd.it/dTB6cXZocXk4aGVlMZ6Ql2t-AEpYbSz_aL0unkgEeC92L3d6cUaq68j-jVoP.png?format=pjpg&auto=webp&s=8fe5ef9ac53acced1d338daa97c0979d98e8def7', 'width': 1774}, 'variants': {}}]} |
||
does anyone know how many billion parameter is deepsek r1 model running on their websites? | 2 | just the title as in v3 i heard 37billion parameter were active | 2025-01-22T05:00:59 | https://www.reddit.com/r/LocalLLaMA/comments/1i73jxw/does_anyone_know_how_many_billion_parameter_is/ | pro_ut3104 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i73jxw | false | null | t3_1i73jxw | /r/LocalLLaMA/comments/1i73jxw/does_anyone_know_how_many_billion_parameter_is/ | false | false | self | 2 | null |
Exploring Local Server for Max 1k Active User Base | 1 | [removed] | 2025-01-22T05:17:36 | https://www.reddit.com/r/LocalLLaMA/comments/1i73ua7/exploring_local_server_for_max_1k_active_user_base/ | Infamous_Jaguar_2151 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i73ua7 | false | null | t3_1i73ua7 | /r/LocalLLaMA/comments/1i73ua7/exploring_local_server_for_max_1k_active_user_base/ | false | false | self | 1 | null |
YOU CAN EXTRACT REASONING FROM R1 AND PASS IT ONTO ANY MODEL | 516 | from @skirano on twitter
By the way, you can extract JUST the reasoning from deepseek-reasoner, which means you can send that thinking process to any model you want before they answer you.
Like here where I turn gpt-3.5 turbo into an absolute genius!
| 2025-01-22T05:22:22 | https://v.redd.it/mbcqadwychee1 | Sensitive-Finger-404 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i73x81 | false | {'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/mbcqadwychee1/DASHPlaylist.mpd?a=1740115356%2CNjEzOTY5NjUwNjk0OTNmZTEzMGU0MmQxZTE3ODNiZDZmYTU0YWEwNGNlZDBjNjAwMTgzZWVkY2UyNTdiMDcyNA%3D%3D&v=1&f=sd', 'duration': 38, 'fallback_url': 'https://v.redd.it/mbcqadwychee1/DASH_720.mp4?source=fallback', 'has_audio': False, 'height': 720, 'hls_url': 'https://v.redd.it/mbcqadwychee1/HLSPlaylist.m3u8?a=1740115356%2CZjc0YTUxNjFiMDY5YmIwMGRiYjgwZTBhNDM4NzIyMjU1MGUxYjVkNTAxYzJiMDdkOGRjM2U5MTkwYmUyNjhiYg%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/mbcqadwychee1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 912}} | t3_1i73x81 | /r/LocalLLaMA/comments/1i73x81/you_can_extract_reasoning_from_r1_and_pass_it/ | false | false | 516 | {'enabled': False, 'images': [{'id': 'OG1uaHRydHljaGVlMeGKc_GKsNSHC_YJy3k1hv6gZ336TNH-m_F1sXruvXhI', 'resolutions': [{'height': 85, 'url': 'https://external-preview.redd.it/OG1uaHRydHljaGVlMeGKc_GKsNSHC_YJy3k1hv6gZ336TNH-m_F1sXruvXhI.png?width=108&crop=smart&format=pjpg&auto=webp&s=0fa874946f0c0c82f9181ff7c5c869a925b1686e', 'width': 108}, {'height': 170, 'url': 'https://external-preview.redd.it/OG1uaHRydHljaGVlMeGKc_GKsNSHC_YJy3k1hv6gZ336TNH-m_F1sXruvXhI.png?width=216&crop=smart&format=pjpg&auto=webp&s=6656dec21d8f45a956ade8e5ca479ffc054c8fed', 'width': 216}, {'height': 252, 'url': 'https://external-preview.redd.it/OG1uaHRydHljaGVlMeGKc_GKsNSHC_YJy3k1hv6gZ336TNH-m_F1sXruvXhI.png?width=320&crop=smart&format=pjpg&auto=webp&s=98b418c492125b7fb9a88bb5448a54bb2fd838e0', 'width': 320}, {'height': 505, 'url': 'https://external-preview.redd.it/OG1uaHRydHljaGVlMeGKc_GKsNSHC_YJy3k1hv6gZ336TNH-m_F1sXruvXhI.png?width=640&crop=smart&format=pjpg&auto=webp&s=802b58e3ae149850ab9276f0303e3d6984444c50', 'width': 640}], 'source': {'height': 720, 'url': 'https://external-preview.redd.it/OG1uaHRydHljaGVlMeGKc_GKsNSHC_YJy3k1hv6gZ336TNH-m_F1sXruvXhI.png?format=pjpg&auto=webp&s=d11b86b01fbc2a87eb165d3a556cb1c4c2bc2fc4', 'width': 912}, 'variants': {}}]} |
|
Deepseek running in my basement finds the 5 odd numbers with the letter e in them | 52 | Amazing. I'm a believer.
It's impossible to find an odd number in English that doesn't have the letter 'e' in its spelling because all single-digit odd numbers (1, 3, 5, 7, 9) contain the letter 'e' in their names. Consequently, any odd number, regardless of its size, will include one of these digits in its ones place, resulting in the presence of the letter 'e'. Therefore, no such numbers exist.
**Answer:** There are no odd numbers in English that don't have the letter 'e' in their spelling.
If you want to see the thinking
[https://pastebin.com/rbvF7p2f](https://pastebin.com/rbvF7p2f) | 2025-01-22T05:24:00 | https://www.reddit.com/r/LocalLLaMA/comments/1i73y5t/deepseek_running_in_my_basement_finds_the_5_odd/ | segmond | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i73y5t | false | null | t3_1i73y5t | /r/LocalLLaMA/comments/1i73y5t/deepseek_running_in_my_basement_finds_the_5_odd/ | false | false | self | 52 | {'enabled': False, 'images': [{'id': 'OgFzGCIRw1ZxjMOSkfV1OiH-_nQiZl8rzSonmOAuhGs', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/P8lS0kk6BFe2IEo6TxCZd1LVwksc34IkzGTVx_SCc8w.jpg?width=108&crop=smart&auto=webp&s=3d74dbe4f1d67cc8b587db9aa01762f26e269bcf', 'width': 108}], 'source': {'height': 150, 'url': 'https://external-preview.redd.it/P8lS0kk6BFe2IEo6TxCZd1LVwksc34IkzGTVx_SCc8w.jpg?auto=webp&s=b9f5c4e4867fbffb2c1ff45dd70aa338d1e3f40c', 'width': 150}, 'variants': {}}]} |
Whats the best OSS alternative to cursor? | 3 | I’ve been using continue on vscode with qwen2.5-coder:7b for code generation and qwen2.5-coder:1.5b for autocomplete, and it hasn’t come close to cursor in terms of functionality. this isn’t just about the models being less effective—the features themselves often don’t work as expected. For example, clicking "apply" on an edited piece of code doesn’t always replace it correctly. there’s also no git-like interface showing green and red differences to highlight changes.
I’m not sure which model would have enough context to chat effectively with the entire codebase. additionally, there’s no swe agent (like composer in cursor). Are there alternative tools that offer smoother functionality? The priority here isn’t necessarily high-quality code output it’s more about seamless usability/functionality. | 2025-01-22T05:31:28 | https://www.reddit.com/r/LocalLLaMA/comments/1i742ha/whats_the_best_oss_alternative_to_cursor/ | CaptTechno | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i742ha | false | null | t3_1i742ha | /r/LocalLLaMA/comments/1i742ha/whats_the_best_oss_alternative_to_cursor/ | false | false | self | 3 | null |
fine tuning Llama on a statistical data | 1 | [removed] | 2025-01-22T05:34:29 | https://www.reddit.com/r/LocalLLaMA/comments/1i7447g/fine_tuning_llama_on_a_statistical_data/ | Aymankoos | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7447g | false | null | t3_1i7447g | /r/LocalLLaMA/comments/1i7447g/fine_tuning_llama_on_a_statistical_data/ | false | false | self | 1 | null |
fine tuning Llama on a statistical data | 1 | [removed] | 2025-01-22T05:35:45 | https://www.reddit.com/r/LocalLLaMA/comments/1i744ym/fine_tuning_llama_on_a_statistical_data/ | Aymankoos | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i744ym | false | null | t3_1i744ym | /r/LocalLLaMA/comments/1i744ym/fine_tuning_llama_on_a_statistical_data/ | false | false | self | 1 | null |
Fine tuning llama 3 on a statistical data | 1 | [removed] | 2025-01-22T05:43:47 | https://www.reddit.com/r/LocalLLaMA/comments/1i749ie/fine_tuning_llama_3_on_a_statistical_data/ | Aymankoos | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i749ie | false | null | t3_1i749ie | /r/LocalLLaMA/comments/1i749ie/fine_tuning_llama_3_on_a_statistical_data/ | false | false | self | 1 | null |
Why my questions automatically removed | 1 | [removed] | 2025-01-22T05:44:50 | https://www.reddit.com/r/LocalLLaMA/comments/1i74a1p/why_my_questions_automatically_removed/ | Aymankoos | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i74a1p | false | null | t3_1i74a1p | /r/LocalLLaMA/comments/1i74a1p/why_my_questions_automatically_removed/ | false | false | self | 1 | null |
BAAI/bge-en-icl not working with OpenWebUI. | 1 | It tries to load the shards and successfully loads 2 shards but then say press any key to continue and when I press a key it just closes:
Loading WEBUI_SECRET_KEY from file, not provided as an environment variable.
Loading WEBUI_SECRET_KEY from H:\AI\OpenWebUI\.webui_secret_key
CUDA is enabled, appending LD_LIBRARY_PATH to include torch/cudnn & cublas libraries.
CUDA seems to be working
H:\AI\OpenWebUI\venv\Lib\site-packages\open_webui
H:\AI\OpenWebUI\venv\Lib\site-packages
H:\AI\OpenWebUI\venv\Lib
Running migrations
INFO [alembic.runtime.migration] Context impl SQLiteImpl.
INFO [alembic.runtime.migration] Will assume non-transactional DDL.
INFO [open_webui.env] 'ENABLE_API_KEY' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_API_KEY_ENDPOINT_RESTRICTIONS' loaded from the latest database entry
INFO [open_webui.env] 'API_KEY_ALLOWED_ENDPOINTS' loaded from the latest database entry
INFO [open_webui.env] 'JWT_EXPIRES_IN' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_OLLAMA_API' loaded from the latest database entry
INFO [open_webui.env] 'OLLAMA_BASE_URLS' loaded from the latest database entry
INFO [open_webui.env] 'OLLAMA_API_CONFIGS' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_OPENAI_API' loaded from the latest database entry
INFO [open_webui.env] 'OPENAI_API_KEYS' loaded from the latest database entry
INFO [open_webui.env] 'OPENAI_API_BASE_URLS' loaded from the latest database entry
INFO [open_webui.env] 'OPENAI_API_CONFIGS' loaded from the latest database entry
INFO [open_webui.env] 'WEBUI_URL' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_SIGNUP' loaded from the latest database entry
INFO [open_webui.env] 'DEFAULT_LOCALE' loaded from the latest database entry
INFO [open_webui.env] 'DEFAULT_PROMPT_SUGGESTIONS' loaded from the latest database entry
INFO [open_webui.env] 'DEFAULT_USER_ROLE' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_CHANNELS' loaded from the latest database entry
INFO [open_webui.env] 'WEBHOOK_URL' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_COMMUNITY_SHARING' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_MESSAGE_RATING' loaded from the latest database entry
WARNI [open_webui.env]
WARNING: CORS_ALLOW_ORIGIN IS SET TO '*' - NOT RECOMMENDED FOR PRODUCTION DEPLOYMENTS.
INFO [open_webui.env] 'SHOW_ADMIN_DETAILS' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_GOOGLE_DRIVE_INTEGRATION' loaded from the latest database entry
INFO [open_webui.env] 'CONTENT_EXTRACTION_ENGINE' loaded from the latest database entry
INFO [open_webui.env] 'TIKA_SERVER_URL' loaded from the latest database entry
INFO [open_webui.env] 'RAG_TOP_K' loaded from the latest database entry
INFO [open_webui.env] 'RAG_RELEVANCE_THRESHOLD' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_RAG_HYBRID_SEARCH' loaded from the latest database entry
INFO [open_webui.env] 'RAG_EMBEDDING_ENGINE' loaded from the latest database entry
INFO [open_webui.env] 'PDF_EXTRACT_IMAGES' loaded from the latest database entry
INFO [open_webui.env] 'RAG_EMBEDDING_MODEL' loaded from the latest database entry
INFO [open_webui.env] Embedding model set: BAAI/bge-en-icl
INFO [open_webui.env] 'RAG_EMBEDDING_BATCH_SIZE' loaded from the latest database entry
INFO [open_webui.env] 'RAG_RERANKING_MODEL' loaded from the latest database entry
INFO [open_webui.env] Reranking model set: mixedbread-ai/mxbai-rerank-large-v1
INFO [open_webui.env] 'RAG_TEXT_SPLITTER' loaded from the latest database entry
INFO [open_webui.env] 'CHUNK_SIZE' loaded from the latest database entry
INFO [open_webui.env] 'CHUNK_OVERLAP' loaded from the latest database entry
INFO [open_webui.env] 'RAG_TEMPLATE' loaded from the latest database entry
INFO [open_webui.env] 'RAG_OPENAI_API_BASE_URL' loaded from the latest database entry
INFO [open_webui.env] 'RAG_OPENAI_API_KEY' loaded from the latest database entry
INFO [open_webui.env] 'RAG_OLLAMA_BASE_URL' loaded from the latest database entry
INFO [open_webui.env] 'RAG_OLLAMA_API_KEY' loaded from the latest database entry
INFO [open_webui.env] 'IMAGE_GENERATION_ENGINE' loaded from the latest database entry
INFO [open_webui.env] 'ENABLE_IMAGE_GENERATION' loaded from the latest database entry
INFO [open_webui.env] 'AUTOMATIC1111_BASE_URL' loaded from the latest database entry
INFO [open_webui.env] 'AUTOMATIC1111_API_AUTH' loaded from the latest database entry
INFO [open_webui.env] 'COMFYUI_BASE_URL' loaded from the latest database entry
INFO [open_webui.env] 'COMFYUI_WORKFLOW' loaded from the latest database entry
INFO [open_webui.env] 'COMFYUI_WORKFLOW' loaded from the latest database entry
INFO [open_webui.env] 'IMAGES_OPENAI_API_BASE_URL' loaded from the latest database entry
INFO [open_webui.env] 'IMAGES_OPENAI_API_KEY' loaded from the latest database entry
WARNI [langchain_community.utils.user_agent] USER_AGENT environment variable not set, consider setting it to identify your requests.
___ __ __ _ _ _ ___
/ _ \ _ __ ___ _ __ \ \ / /__| |__ | | | |_ _|
| | | | '_ \ / _ \ '_ \ \ \ /\ / / _ \ '_ \| | | || |
| |_| | |_) | __/ | | | \ V V / __/ |_) | |_| || |
\___/| .__/ \___|_| |_| \_/\_/ \___|_.__/ \___/|___|
|_|
v0.5.4 - building the best open-source AI user interface.
https://github.com/open-webui/open-webui
Fetching 17 files: 100%|███████████████████████████████████████████████████████████████████████| 17/17 [00:00<?, ?it/s]
Loading checkpoint shards: 33%|███████████████████ | 1/3 [00:02<00:04, 2.30s/it]
Press any key to continue . . . | 2025-01-22T05:58:22 | https://www.reddit.com/r/LocalLLaMA/comments/1i74ho3/baaibgeenicl_not_working_with_openwebui/ | AlgorithmicKing | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i74ho3 | false | null | t3_1i74ho3 | /r/LocalLLaMA/comments/1i74ho3/baaibgeenicl_not_working_with_openwebui/ | false | false | self | 1 | null |
Exploring Local Server for Max 1k Active User Base | 2 |
Hi asking for a friend,
If one wanted to serve a 32b model for a local application to around 1k max (likely 500) monthly active users what would be a good and cost effective option?
Could he put together an epyc server with some a6000s? Having difficulty understanding how the hardware needs to scale for this in terms of vram, ram and cpu compute power.
He would essentially be hosting an ai application. Hoping to save over cloud computing costs. | 2025-01-22T06:00:57 | https://www.reddit.com/r/LocalLLaMA/comments/1i74j9k/exploring_local_server_for_max_1k_active_user_base/ | NewBronzeAge | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i74j9k | false | null | t3_1i74j9k | /r/LocalLLaMA/comments/1i74j9k/exploring_local_server_for_max_1k_active_user_base/ | false | false | self | 2 | null |
DeepSeek R1 32B is way better than 7B Distill, even at Q4 quant | 51 | I've been quite impressed by the model. I'm using the Qwen distill and so far it's working well, although as is typical with these models, they tend to overthink a lot! But it answered my trick question in one shot (See comments). | 2025-01-22T06:01:35 | https://www.reddit.com/r/LocalLLaMA/comments/1i74jqg/deepseek_r1_32b_is_way_better_than_7b_distill/ | nderstand2grow | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i74jqg | false | null | t3_1i74jqg | /r/LocalLLaMA/comments/1i74jqg/deepseek_r1_32b_is_way_better_than_7b_distill/ | false | false | self | 51 | null |
How trusted is LM Studio? | 7 | I apologize if this isn't the place to ask and will delete the post if not. I'm an absolute beginner with this stuff, so i apologize for my ignorance. I had seen LM studio was a beginner friendly intro to the world of local LLM and wanted to try it out. My more experienced friend said that I shouldn't bother with it and not to trust it. Is it something that the community stands by or is it not trustworthy? | 2025-01-22T06:06:05 | https://www.reddit.com/r/LocalLLaMA/comments/1i74mao/how_trusted_is_lm_studio/ | DevilBirb | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i74mao | false | null | t3_1i74mao | /r/LocalLLaMA/comments/1i74mao/how_trusted_is_lm_studio/ | false | false | self | 7 | null |
M4 Max 16-Core, 40-Core Graphics - 128 GB Ram | 1 | [removed] | 2025-01-22T06:33:42 | https://www.reddit.com/r/LocalLLaMA/comments/1i7514d/m4_max_16core_40core_graphics_128_gb_ram/ | Vect0r_YT | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7514d | false | null | t3_1i7514d | /r/LocalLLaMA/comments/1i7514d/m4_max_16core_40core_graphics_128_gb_ram/ | false | false | self | 1 | null |
How to debug eval metrics? [See description] | 1 | Hi All,
I am looking to host an offline/local solution to view/interpret the standard-eval outputs from different LLMs. Is there something I can use locally?
I have the outputs in a local jsonl file, but I want some locally-hosted frontend which takes in the filename and then gives an easy way to play around with the outputs. Having metadata like average len of inputs, avg output tokens etc would also be useful. Any pointers?
Thanks. | 2025-01-22T06:34:00 | https://www.reddit.com/r/LocalLLaMA/comments/1i751a2/how_to_debug_eval_metrics_see_description/ | weight_matrix | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i751a2 | false | null | t3_1i751a2 | /r/LocalLLaMA/comments/1i751a2/how_to_debug_eval_metrics_see_description/ | false | false | self | 1 | null |
generating open-source synthetic datasets for data-scarce tasks | 1 | [removed] | 2025-01-22T06:36:38 | https://www.reddit.com/r/LocalLLaMA/comments/1i752n9/generating_opensource_synthetic_datasets_for/ | Classic_Eggplant8827 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i752n9 | false | null | t3_1i752n9 | /r/LocalLLaMA/comments/1i752n9/generating_opensource_synthetic_datasets_for/ | false | false | self | 1 | null |
Best frameworks for fine-tuning models—what’s everyone using? | 1 | [removed] | 2025-01-22T06:38:14 | https://www.reddit.com/r/LocalLLaMA/comments/1i753he/best_frameworks_for_finetuning_modelswhats/ | Vivid-Entertainer752 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i753he | false | null | t3_1i753he | /r/LocalLLaMA/comments/1i753he/best_frameworks_for_finetuning_modelswhats/ | false | false | self | 1 | null |
I don’t believe the $500 AI investment | 1 | Looking at this deal, several things don't add up. The $500 billion figure is wildly optimistic - that's almost double what the entire US government committed to semiconductor manufacturing through the CHIPS Act. When you dig deeper, you see lots of vague promises but no real details about where the money's coming from or how they'll actually build anything.
The legal language is especially fishy. Instead of making firm commitments, they're using weasel words like "intends to," "evaluating," and "potential partnerships." This isn't accidental - by running everything through Stargate, a new private company, and using this careful language, they've created a perfect shield for bigger players like SoftBank and Microsoft. If things go south, they can just blame "market conditions" and walk away with minimal exposure. Private companies like Stargate don't face the same strict disclosure requirements as public ones.
The timing is also telling - announcing this massive investment right after Trump won the presidency was clearly designed for maximum political impact. It fits perfectly into the narrative of bringing jobs and investment back to America. Using inflated job numbers for data centers (which typically employ relatively few people once built) while making vague promises about US technological leadership? That’s politics.
My guess? There's probably a real data center project in the works, but it's being massively oversold for publicity and political gains. The actual investment will likely be much smaller, take longer to complete, and involve different partners than what's being claimed. This announcement just is a deal structured by lawyers who wanted to generate maximum headlines while minimizing any legal risk for their clients. | 2025-01-22T07:01:24 | https://www.reddit.com/r/LocalLLaMA/comments/1i75fde/i_dont_believe_the_500_ai_investment/ | MattDTO | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i75fde | false | null | t3_1i75fde | /r/LocalLLaMA/comments/1i75fde/i_dont_believe_the_500_ai_investment/ | false | false | self | 1 | null |
I (a content writer) built a Shopify store using AI code generators | 0 | The process was a lot simpler than you'd expect. I signed up for Shopify, chose a plan, generated code snippets using AI (this was the most complex part, I used HTML, CSS, and JS for the backend), and added them to Shopify's theme editor. It displayed errors for certain things (like Google fonts) but those I was able to fix with AI as well.
I wrote a detailed showcase of the process (with screenshots) for anyone interested: [https://blog.getbind.co/2025/01/21/how-to-create-a-shopify-store-with-ai/](https://blog.getbind.co/2025/01/21/how-to-create-a-shopify-store-with-ai/)
**NOTE:** this was only a project and not a real venture. | 2025-01-22T07:02:20 | https://www.reddit.com/r/LocalLLaMA/comments/1i75fvr/i_a_content_writer_built_a_shopify_store_using_ai/ | johnzakma10 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i75fvr | false | null | t3_1i75fvr | /r/LocalLLaMA/comments/1i75fvr/i_a_content_writer_built_a_shopify_store_using_ai/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': '1_MnsoBOjHUVlBv8s1AW8GF3ZoHqy4Q7Cx8Vh-5po64', 'resolutions': [{'height': 23, 'url': 'https://external-preview.redd.it/qucmYrGLR_ezD9eMXgpepPDC5n8MtQ-JNXioY_ynCHg.jpg?width=108&crop=smart&auto=webp&s=5d3f084b1f24c6be1b219ed06d50ede11039ae20', 'width': 108}, {'height': 47, 'url': 'https://external-preview.redd.it/qucmYrGLR_ezD9eMXgpepPDC5n8MtQ-JNXioY_ynCHg.jpg?width=216&crop=smart&auto=webp&s=4c4c27a0375b804db5d90bf12bf5c57a81b64386', 'width': 216}], 'source': {'height': 60, 'url': 'https://external-preview.redd.it/qucmYrGLR_ezD9eMXgpepPDC5n8MtQ-JNXioY_ynCHg.jpg?auto=webp&s=58df702c38afd9cce5d0d8f1b6181031aa15e77b', 'width': 272}, 'variants': {}}]} |
I don’t believe the $500 Billion OpenAI investment | 547 | Looking at this deal, several things don't add up. The $500 billion figure is wildly optimistic - that's almost double what the entire US government committed to semiconductor manufacturing through the CHIPS Act. When you dig deeper, you see lots of vague promises but no real details about where the money's coming from or how they'll actually build anything.
The legal language is especially fishy. Instead of making firm commitments, they're using weasel words like "intends to," "evaluating," and "potential partnerships." This isn't accidental - by running everything through Stargate, a new private company, and using this careful language, they've created a perfect shield for bigger players like SoftBank and Microsoft. If things go south, they can just blame "market conditions" and walk away with minimal exposure. Private companies like Stargate don't face the same strict disclosure requirements as public ones.
The timing is also telling - announcing this massive investment right after Trump won the presidency was clearly designed for maximum political impact. It fits perfectly into the narrative of bringing jobs and investment back to America. Using inflated job numbers for data centers (which typically employ relatively few people once built) while making vague promises about US technological leadership? That’s politics.
My guess? There's probably a real data center project in the works, but it's being massively oversold for publicity and political gains. The actual investment will likely be much smaller, take longer to complete, and involve different partners than what's being claimed. This announcement just is a deal structured by lawyers who wanted to generate maximum headlines while minimizing any legal risk for their clients. | 2025-01-22T07:03:02 | https://www.reddit.com/r/LocalLLaMA/comments/1i75g7p/i_dont_believe_the_500_billion_openai_investment/ | MattDTO | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i75g7p | false | null | t3_1i75g7p | /r/LocalLLaMA/comments/1i75g7p/i_dont_believe_the_500_billion_openai_investment/ | false | false | self | 547 | null |
deepseek-r1:14b - attempting to answer version differences on qt with bit tricky questions | 4 | Have not expected, at least local model's give reasonable answers so soon. I am running deepseek-r1:14b on RTX3060. every time, when I try with local models not satisfied with answers after couple of tries. So switching back to LLM's like openai or claude or gemini. Now I see we can start using local models reasonably. though answers are not exact, but the reasoning model like this, and showing thinking in the direction, I see feel comfortable to use.
https://preview.redd.it/772ekq3ruhee1.png?width=3480&format=png&auto=webp&s=db2ea99d3fc168372f0b9b1e48f6aae2e002fe91
| 2025-01-22T07:07:46 | https://www.reddit.com/r/LocalLLaMA/comments/1i75iny/deepseekr114b_attempting_to_answer_version/ | prabhic | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i75iny | false | null | t3_1i75iny | /r/LocalLLaMA/comments/1i75iny/deepseekr114b_attempting_to_answer_version/ | false | false | 4 | null |
|
AI News Recommendations? | 6 | Hello! I'm looking for recommendations for AI news sources, similar to what I currently listen to. I regularly follow these audio/video sources:
* Latent Space: The AI Engineer Podcast
* Machine Learning Street Talk
* aiexplained-official (YouTube)
* MattVidPro (YouTube)
I download the audio from these videos each morning using Newpipe and listen to them in Antennapod. I also read the Buttondown AI newsletter (recommended by karpathy) and wish there was an audio version of that too.
Based on my current sources, what other AI news sources would you recommend, especially audio-based ones? | 2025-01-22T07:14:21 | https://www.reddit.com/r/LocalLLaMA/comments/1i75m1u/ai_news_recommendations/ | dewijones92 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i75m1u | false | null | t3_1i75m1u | /r/LocalLLaMA/comments/1i75m1u/ai_news_recommendations/ | false | false | self | 6 | null |
[R] giving out gpus for free to make any niche specialized model | 8 | hello so the story was that i raised money from vcs to start a company - i actually hate this (i'm extremely unhappy). i decided that what would make me extremely happy is to fund a community research effort around what is possible with synthetic data. i am giving gpus to people to generate cool function-calling or multi-turn datasets and populate huggingface with really cool niche sota agents in tasks where llms are notoriously bad in (rust, text2sql, verilog, etc). if you have something you want to work on dm me! | 2025-01-22T07:25:47 | https://www.reddit.com/r/LocalLLaMA/comments/1i75rnq/r_giving_out_gpus_for_free_to_make_any_niche/ | Classic_Eggplant8827 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i75rnq | false | null | t3_1i75rnq | /r/LocalLLaMA/comments/1i75rnq/r_giving_out_gpus_for_free_to_make_any_niche/ | false | false | self | 8 | null |
Bytedance Model: Doubao 1.5pro - Doubao Team | 21 | 2025-01-22T07:36:36 | https://team.doubao.com/en/special/doubao_1_5_pro | a445141126 | team.doubao.com | 1970-01-01T00:00:00 | 0 | {} | 1i75wyo | false | null | t3_1i75wyo | /r/LocalLLaMA/comments/1i75wyo/bytedance_model_doubao_15pro_doubao_team/ | false | false | default | 21 | null |
|
Cloud hosted LLaMA or ChatGPT Embeddings | 1 | [removed] | 2025-01-22T07:47:39 | https://www.reddit.com/r/LocalLLaMA/comments/1i762ec/cloud_hosted_llama_or_chatgpt_embeddings/ | Extension-Lettuce623 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i762ec | false | null | t3_1i762ec | /r/LocalLLaMA/comments/1i762ec/cloud_hosted_llama_or_chatgpt_embeddings/ | false | false | self | 1 | null |
R1-Zero: Pure RL Creates a Mind We Can’t Decode—Is This AGI’s Dark Mirror? | 261 | The AI world is losing its mind over DeepSeek-R1-Zero, a model that skipped supervised fine-tuning (SFT) entirely and learned purely through reinforcement learning (RL). Unlike its sibling R1—which uses **some** SFT data to stay "human-readable"—R1-Zero’s training mirrors AlphaZero’s trial-and-error self-play. The result? **Jaw-dropping performance** (AIME math scores jumped from 15.6% → 86.7%) paired with **bizarre, uninterpretable reasoning**. Researchers observed "aha moments" where it autonomously rechecked flawed logic mid-process and allocated more compute to harder problems—**without human guidance**. But here’s the kicker: its outputs are riddled with garbled language mixes (e.g., Chinese/English spaghetti code) and logic leaps that even its creators can’t fully explain.
Meanwhile, R1 (the SFT-hybrid version) achieves similar performance **without the chaos**, proving that human-curated data still tames the beast. But at what cost? R1-Zero’s pure RL approach hints at a terrifying possibility: **minds that optimize truth beyond human comprehension**. And with API costs 50x cheaper than OpenAI’s, scaling this could democratize superintelligence—or unleash unreadable black-box AI.
If R1-Zero’s "alien logic" solves problems we can’t, does readability even matter… or is this how alignment dies? | 2025-01-22T07:54:24 | https://www.reddit.com/r/LocalLLaMA/comments/1i765q0/r1zero_pure_rl_creates_a_mind_we_cant_decodeis/ | Fun_Dragonfruit_4613 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i765q0 | false | null | t3_1i765q0 | /r/LocalLLaMA/comments/1i765q0/r1zero_pure_rl_creates_a_mind_we_cant_decodeis/ | false | false | self | 261 | null |
rStar-Math? | 2 | Here is rStar-Math. Has there been any exploration of its integration with Deepseek distiled models? I am interested in obtaining an informed critique of such undertakings.
[https://github.com/microsoft/rStar](https://github.com/microsoft/rStar) | 2025-01-22T08:11:52 | https://www.reddit.com/r/LocalLLaMA/comments/1i76e1z/rstarmath/ | Wonderful_Second5322 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i76e1z | false | null | t3_1i76e1z | /r/LocalLLaMA/comments/1i76e1z/rstarmath/ | false | false | self | 2 | {'enabled': False, 'images': [{'id': 'a-XdqHvGm-BQEECrtM5DxV5UrUOx6307m1WuUZqPmUM', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/EBkeap0N69awf8OnUeRy8W1_ZIhMW7VTuxrBYV1pWrE.jpg?width=108&crop=smart&auto=webp&s=dde75b754ab5f0783c3236039d8e22a82739779c', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/EBkeap0N69awf8OnUeRy8W1_ZIhMW7VTuxrBYV1pWrE.jpg?width=216&crop=smart&auto=webp&s=d363e7641caedebf49add2434c45d7459e218ed1', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/EBkeap0N69awf8OnUeRy8W1_ZIhMW7VTuxrBYV1pWrE.jpg?width=320&crop=smart&auto=webp&s=9ff7e2cdb1b25e1c23b1707773d9e2798d0dbd65', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/EBkeap0N69awf8OnUeRy8W1_ZIhMW7VTuxrBYV1pWrE.jpg?width=640&crop=smart&auto=webp&s=f194fbd19a9e4e0251c01c3241b6bdcabef718d7', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/EBkeap0N69awf8OnUeRy8W1_ZIhMW7VTuxrBYV1pWrE.jpg?width=960&crop=smart&auto=webp&s=755d0f2d78097e1a63540bcc8da971eea1700203', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/EBkeap0N69awf8OnUeRy8W1_ZIhMW7VTuxrBYV1pWrE.jpg?width=1080&crop=smart&auto=webp&s=f179568a7e93a3de77bdba9ce6dd9ba63b2a971e', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/EBkeap0N69awf8OnUeRy8W1_ZIhMW7VTuxrBYV1pWrE.jpg?auto=webp&s=c43400325bd3b5c598650256882db93d84387bd2', 'width': 1200}, 'variants': {}}]} |
LoRA for LLM? | 1 | [removed] | 2025-01-22T08:15:41 | https://www.reddit.com/r/LocalLLaMA/comments/1i76frg/lora_for_llm/ | berrj0108 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i76frg | false | null | t3_1i76frg | /r/LocalLLaMA/comments/1i76frg/lora_for_llm/ | false | false | self | 1 | null |
Is the Deepseek R1 training code available anywhere? | 2 | I tried searching online but everywhere I look its just the code to run inference rather than the complete code. I would like to know what they did and how it works, sort off reverse engineer it but I cannot find the code anywhere. | 2025-01-22T08:43:52 | https://www.reddit.com/r/LocalLLaMA/comments/1i76sjb/is_the_deepseek_r1_training_code_available/ | riansar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i76sjb | false | null | t3_1i76sjb | /r/LocalLLaMA/comments/1i76sjb/is_the_deepseek_r1_training_code_available/ | false | false | self | 2 | null |
Test `DeepSeek R1 70B` for free using private Fluid API | 8 | 2025-01-22T08:53:55 | https://v.redd.it/721oc32ceiee1 | PrivacyIsImportan1 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i76x7u | false | {'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/721oc32ceiee1/DASHPlaylist.mpd?a=1740128049%2CZDc3MGMzNDRkM2YyYjUyNGU4YzdkNTg3ZGNhZDkwMDcxOGFkYzIyNGUzN2ExZDNkMzZmNzFlZWU2NjUzMWFjMA%3D%3D&v=1&f=sd', 'duration': 19, 'fallback_url': 'https://v.redd.it/721oc32ceiee1/DASH_720.mp4?source=fallback', 'has_audio': False, 'height': 720, 'hls_url': 'https://v.redd.it/721oc32ceiee1/HLSPlaylist.m3u8?a=1740128049%2CODE0NmYxY2U0YTdjNmVlZDA3ZjQ1NzJjMWM1YTcyMjIyMGEwMDc0NmJhNzJlODBmYzhkN2IzMWJkMTZjMzM4OA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/721oc32ceiee1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1176}} | t3_1i76x7u | /r/LocalLLaMA/comments/1i76x7u/test_deepseek_r1_70b_for_free_using_private_fluid/ | false | false | 8 | {'enabled': False, 'images': [{'id': 'dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0', 'resolutions': [{'height': 66, 'url': 'https://external-preview.redd.it/dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0.png?width=108&crop=smart&format=pjpg&auto=webp&s=0dd446924cfbf8002469bedce5c11a6d6a01327e', 'width': 108}, {'height': 132, 'url': 'https://external-preview.redd.it/dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0.png?width=216&crop=smart&format=pjpg&auto=webp&s=9b071776d79df9c59d745914925eb28b1b9ef69b', 'width': 216}, {'height': 195, 'url': 'https://external-preview.redd.it/dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0.png?width=320&crop=smart&format=pjpg&auto=webp&s=aa5e0091ed95aeb8d51dbb21430a7598063caa90', 'width': 320}, {'height': 391, 'url': 'https://external-preview.redd.it/dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0.png?width=640&crop=smart&format=pjpg&auto=webp&s=396282c142fd182b4a90923a0f7ada448249496b', 'width': 640}, {'height': 587, 'url': 'https://external-preview.redd.it/dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0.png?width=960&crop=smart&format=pjpg&auto=webp&s=40fefafcad52c85a3e906a48f3e8189ed35d3459', 'width': 960}, {'height': 661, 'url': 'https://external-preview.redd.it/dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0.png?width=1080&crop=smart&format=pjpg&auto=webp&s=5b14969013aca01041a5550d59bd5f5bf0f1b3c9', 'width': 1080}], 'source': {'height': 982, 'url': 'https://external-preview.redd.it/dW55ajc3dmZlaWVlMdx5GDc3_ez7Z-Gp5NX8lFb_ntCghspMTNPuJiKJAyH0.png?format=pjpg&auto=webp&s=238e30a59a15c44152dac3cf18bcc8fd04dfc0cb', 'width': 1604}, 'variants': {}}]} |
||
Built an AI Flow for Client Research and Insights | 1 | I created an AI workflow to automate client research, quickly gathering and analysing insights about potential clients. Here's the process:
**Step 1: Search for the prospect**
Fetch the prospect's professional background, role, and recent activities using a search block. Takes in Name, Company Details as input.
**Step 2: Search company details**
Gather company information like industry, market position, and key data.
**Step 3: Crawl the company website**
Extract details about products, solutions, and news via a web crawl block.
**Step 4: Synthesize insights**
Use an LLM call block to analyze and summarize findings, highlighting opportunities and pain points.
**Step 5: Email the summary**
Automatically send the synthesized report via email.
**You can copy my flow, try it out for free and build yours here:** [https://app.athina.ai/flows/templates/ee41f8e0-5958-45bb-ad82-b60f0e69320d](https://app.athina.ai/flows/templates/ee41f8e0-5958-45bb-ad82-b60f0e69320d)
**Output:**
Tried this for a friend and here are the results:
https://preview.redd.it/wedemg4vfiee1.png?width=1494&format=png&auto=webp&s=ce972327a5661b9b238eebab84fc651baff2b8c0
| 2025-01-22T09:02:04 | https://www.reddit.com/r/LocalLLaMA/comments/1i771a0/built_an_ai_flow_for_client_research_and_insights/ | Sam_Tech1 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i771a0 | false | null | t3_1i771a0 | /r/LocalLLaMA/comments/1i771a0/built_an_ai_flow_for_client_research_and_insights/ | false | false | 1 | null |
|
Alright, so R1 is great, but have you noticed this? | 1 | [removed] | 2025-01-22T09:04:20 | https://www.reddit.com/r/LocalLLaMA/comments/1i772bf/alright_so_r1_is_great_but_have_you_noticed_this/ | culoacido69420 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i772bf | false | null | t3_1i772bf | /r/LocalLLaMA/comments/1i772bf/alright_so_r1_is_great_but_have_you_noticed_this/ | false | false | self | 1 | null |
Is Deepseek an OpenAI product ?? It says yes !! | 0 | I asked DeepSeek, on its own platform, about the existence of limits on the number of messages, tokens, etc... and it answered me as if it were Open AI:
“Check the OpenAI Help or Support page for up-to-date details, as limits may be adjusted periodically.”
I then asked it:
“but... are you openai?” and it answered me:
“Ah, good question! I am a language model based on OpenAI technology (like GPT-4), but I am not the OpenAI company itself....
... In short: I'm an OpenAI product, but I don't represent the company in real time or manage its policies. was that clear to you? 😊”
???? | 2025-01-22T09:11:47 | https://www.reddit.com/r/LocalLLaMA/comments/1i775uq/is_deepseek_an_openai_product_it_says_yes/ | Epictetito | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i775uq | false | null | t3_1i775uq | /r/LocalLLaMA/comments/1i775uq/is_deepseek_an_openai_product_it_says_yes/ | false | false | self | 0 | null |
Rust code generation with small models | 1 | Can any of the small models generate consistent rust code at least for simple tasks?
Cant find any benchmarks for rust code generation for small models between like 7 to 13b params
Trying codegemma 7b and codellama 13b but they are kinda bad, and just want to talk about implementing features without writing any compilable code, even though I prompt them to do that. Im using 3060, so VRAM is limited to 12 gb.
Any advice would be greatly appreciated | 2025-01-22T09:15:47 | https://www.reddit.com/r/LocalLLaMA/comments/1i777pw/rust_code_generation_with_small_models/ | ChazychazZz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i777pw | false | null | t3_1i777pw | /r/LocalLLaMA/comments/1i777pw/rust_code_generation_with_small_models/ | false | false | self | 1 | null |
How to Optimize GPU Utilization Across Multiple vLLM Docker Instances on a 4-GPU Server? | 1 | [removed] | 2025-01-22T09:16:05 | https://www.reddit.com/r/LocalLLaMA/comments/1i777uf/how_to_optimize_gpu_utilization_across_multiple/ | EstimateFearless4742 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i777uf | false | null | t3_1i777uf | /r/LocalLLaMA/comments/1i777uf/how_to_optimize_gpu_utilization_across_multiple/ | false | false | self | 1 | null |
Need Advice: Building a Local RAG-Based Chatbot for a 1000-User Company - OpenWebUI vs. AnythingLLM? | 0 | I'm planning to build a local RAG (Retrieval-Augmented Generation)-based chatbot for a 1000-user company, and I'm stuck between using OpenWebUI and AnythingLLM.
Here's what I'm aiming for:
1. Implementing custom tools, pipelines, and API integrations.
2. Experimenting with different RAG approaches, including LightRAG.
3. Managing user roles, permissions, and groupings effectively.
Given these requirements, which project would be a better investment? I want to ensure scalability, flexibility, and a clear development path. Any insights, experiences, or recommendations would be greatly appreciated! | 2025-01-22T09:21:07 | https://www.reddit.com/r/LocalLLaMA/comments/1i77a2o/need_advice_building_a_local_ragbased_chatbot_for/ | PickkNickk | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i77a2o | false | null | t3_1i77a2o | /r/LocalLLaMA/comments/1i77a2o/need_advice_building_a_local_ragbased_chatbot_for/ | false | false | self | 0 | null |
Stubborn Strawberry Deepseek R1 14b | 1 | deepseek-r1:14b 8:20 PM
<think> Alright, the user is pointing out specific positions for 'r's in "strawberry," claiming there are three.
I need to carefully analyze the word again, perhaps assigning numerical indices to each letter.
Wait, if we number letters starting from 1:
1. S
2. T
3. R
4. A
5. W
6. B
7. E
8. R
9. R
10. Y
Here, 'R's are at positions 3, 8, and 9.
Wait a second—that means there are three 'r's? But I thought it was two.
I must have made a mistake in my earlier analysis.
Hold on, the correct spelling of "strawberry" is **S-T-R-A-W-B-E-R-R-Y** , which indeed has two 'R's, not three. The confusion might arise from how we count positions.
Let me clarify this to avoid further misunderstanding. </think>
Ah, I see where you're coming from! Let’s assign numerical indices to each letter in the word "strawberry":
1. S
2. T
3. R
4. A
5. W
6. B
7. E
8. R
9. R
10. Y
As shown above, there are **two "R"s** in positions 3 and 9. There is no third "R" in the word.
If you're counting three "R"s, it might be due to a misunderstanding of how letters are positioned or a different spelling variation. Let me know if you'd like further clarification! 😊 | 2025-01-22T09:27:38 | https://www.reddit.com/r/LocalLLaMA/comments/1i77d0c/stubborn_strawberry_deepseek_r1_14b/ | Rogue_NPC | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i77d0c | false | null | t3_1i77d0c | /r/LocalLLaMA/comments/1i77d0c/stubborn_strawberry_deepseek_r1_14b/ | false | false | self | 1 | null |
Exploring Advanced RAG Techniques: Are Current Solutions Too Naive? | 10 | In projects like AnythingLLM, Dify, and OpenWebUI, I often find their RAG (Retrieval-Augmented Generation) techniques to be quite naive. While they get the job done for simple use cases, I’m curious about the potential for implementing more complex and effective RAG approaches.
For example:
What are the best practices for achieving higher retrieval precision and context relevancy?
Are there open-source frameworks or tools that allow for more advanced pipeline customization?
How can we go beyond the basics and implement state-of-the-art RAG techniques effectively?
I’m particularly interested in hearing about projects or ideas that push the boundaries of RAG implementation. Any insights or resources on this topic would be greatly appreciated!
| 2025-01-22T09:28:09 | https://www.reddit.com/r/LocalLLaMA/comments/1i77d8t/exploring_advanced_rag_techniques_are_current/ | PickkNickk | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i77d8t | false | null | t3_1i77d8t | /r/LocalLLaMA/comments/1i77d8t/exploring_advanced_rag_techniques_are_current/ | false | false | self | 10 | null |
Best frameworks for fine-tuning models—what’s everyone using? | 1 | [removed] | 2025-01-22T09:33:32 | https://www.reddit.com/r/LocalLLaMA/comments/1i77foq/best_frameworks_for_finetuning_modelswhats/ | Vivid-Entertainer752 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i77foq | false | null | t3_1i77foq | /r/LocalLLaMA/comments/1i77foq/best_frameworks_for_finetuning_modelswhats/ | false | false | self | 1 | null |
Test `DeepSeek R1 70B` for free using private Fluid API | 2 | Hi, someone asked where to test DeepSeek R1 Llama 70B, so I thought I'll share a few free accounts here:
Endpoint: # OpenAI compatible (partial support)
API Keys: 5022973707673231 # Authorization: Bearer {account_no}
3272534560577170
4578564232362932
5156468097321794
7111624403950538
1791875353587484
5678862424161157
Model name: Fluid-Latest
(Please don't overuse them)https://api.getfluid.app/v1
You can also use my app [https://getfluid.app](https://getfluid.app) as a client if you have a Mac. (Here's [how to unlock](https://changelog.getfluid.app/#demo) Pro)
If you are seeking for a private API provider with a strict privacy policy (no logging, no persistent storage, no sharing, hosted in EU) - check out [details here](https://getfluid.app/#fluidpro). Use `LOCAL_LLAMA` code to get 50% perpetual discount (30 codes available).
\--------------------------------------------------
So far I'm very impressed with R1 70B, although it's clearly limited by the underlying Llama 3.3. For many everyday tasks it's noticeably better, sometimes a bit disappointing (e.g. coding in SwiftUI). I tested both FP4 (AWQ) and FP8 quants, and didn't notice much difference (maybe had wrong FP8 quant?), so went with AWQ for its speed. Fun thing, within Fluid app I switched to local Llama 8b to continue R1 conversation, and it immediately started thinking in the same way as R1, using <think> tags - I thought I had a bug for a moment.
(note: reposting this, as my previous post didn't display the comment with api keys for some reason). | 2025-01-22T09:47:07 | https://www.reddit.com/r/LocalLLaMA/comments/1i77m08/test_deepseek_r1_70b_for_free_using_private_fluid/ | PrivacyIsImportan1 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i77m08 | false | null | t3_1i77m08 | /r/LocalLLaMA/comments/1i77m08/test_deepseek_r1_70b_for_free_using_private_fluid/ | false | false | self | 2 | {'enabled': False, 'images': [{'id': 'dvoD3QHOC6zmJFTNoOo7-Do6OcHwF_xlrSRYbkSnzMs', 'resolutions': [{'height': 64, 'url': 'https://external-preview.redd.it/JT6-SrnAvvvG9Bbs6RHeqrEkBhNTjNdD89aMIRASVhE.jpg?width=108&crop=smart&auto=webp&s=969f4ab96196b3ef88a01b12e4f55e2fe4be33c3', 'width': 108}, {'height': 128, 'url': 'https://external-preview.redd.it/JT6-SrnAvvvG9Bbs6RHeqrEkBhNTjNdD89aMIRASVhE.jpg?width=216&crop=smart&auto=webp&s=ff6a54f853352312da6266a017473c470ca536fc', 'width': 216}, {'height': 190, 'url': 'https://external-preview.redd.it/JT6-SrnAvvvG9Bbs6RHeqrEkBhNTjNdD89aMIRASVhE.jpg?width=320&crop=smart&auto=webp&s=b37e764a9cc46b402299fb32ad6e13ae881ccc58', 'width': 320}, {'height': 380, 'url': 'https://external-preview.redd.it/JT6-SrnAvvvG9Bbs6RHeqrEkBhNTjNdD89aMIRASVhE.jpg?width=640&crop=smart&auto=webp&s=71704f304eb18d08a695c22dd0743e67b6a9da02', 'width': 640}, {'height': 571, 'url': 'https://external-preview.redd.it/JT6-SrnAvvvG9Bbs6RHeqrEkBhNTjNdD89aMIRASVhE.jpg?width=960&crop=smart&auto=webp&s=37e7d56f0e644629ccca9527cac4ac96513315a5', 'width': 960}, {'height': 642, 'url': 'https://external-preview.redd.it/JT6-SrnAvvvG9Bbs6RHeqrEkBhNTjNdD89aMIRASVhE.jpg?width=1080&crop=smart&auto=webp&s=37b6348b021a964465669248e7e9da9f72fae23b', 'width': 1080}], 'source': {'height': 714, 'url': 'https://external-preview.redd.it/JT6-SrnAvvvG9Bbs6RHeqrEkBhNTjNdD89aMIRASVhE.jpg?auto=webp&s=46dbd09fb294bf736d07eeaedfee39e549264497', 'width': 1200}, 'variants': {}}]} |
How it feels... | 788 | 2025-01-22T10:01:42 | TheLogiqueViper | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i77sy1 | false | null | t3_1i77sy1 | /r/LocalLLaMA/comments/1i77sy1/how_it_feels/ | false | false | 788 | {'enabled': True, 'images': [{'id': 'nb1tugP6mUBF75kaTCDNv9yEMzdBeCXYQHzuSCPjxrA', 'resolutions': [{'height': 161, 'url': 'https://preview.redd.it/rocl1zwsqiee1.png?width=108&crop=smart&auto=webp&s=2b707f5a9aff527803097536e169098785f9c768', 'width': 108}, {'height': 323, 'url': 'https://preview.redd.it/rocl1zwsqiee1.png?width=216&crop=smart&auto=webp&s=e827be10d1a8fbd1bd0f91e73ccb0b8a30467f40', 'width': 216}, {'height': 479, 'url': 'https://preview.redd.it/rocl1zwsqiee1.png?width=320&crop=smart&auto=webp&s=3e2760d53f037795375c623661e55d86333f5bbd', 'width': 320}, {'height': 958, 'url': 'https://preview.redd.it/rocl1zwsqiee1.png?width=640&crop=smart&auto=webp&s=0cf91de1f8277f67834bb8422547f586738360e5', 'width': 640}, {'height': 1438, 'url': 'https://preview.redd.it/rocl1zwsqiee1.png?width=960&crop=smart&auto=webp&s=041a779739d055fa5dcaa70653dc78a257f261d5', 'width': 960}, {'height': 1618, 'url': 'https://preview.redd.it/rocl1zwsqiee1.png?width=1080&crop=smart&auto=webp&s=af7a5d7ae08d81e0480c67420431c83089227bb2', 'width': 1080}], 'source': {'height': 1618, 'url': 'https://preview.redd.it/rocl1zwsqiee1.png?auto=webp&s=254f3a68e602e79ed9faa857e230008721e52b90', 'width': 1080}, 'variants': {}}]} |
|||
How to train an LLM | 1 | [removed] | 2025-01-22T10:12:09 | https://www.reddit.com/r/LocalLLaMA/comments/1i77y3b/how_to_train_an_llm/ | Asleep_Cartoonist460 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i77y3b | false | null | t3_1i77y3b | /r/LocalLLaMA/comments/1i77y3b/how_to_train_an_llm/ | false | false | self | 1 | null |
Stargate - Any reason to be optimistic here for local models..? Or is it just a boost in the arm for ClosedAI? | 1 | [removed] | 2025-01-22T10:20:12 | https://www.reddit.com/r/LocalLLaMA/comments/1i7823a/stargate_any_reason_to_be_optimistic_here_for/ | mark-lord | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7823a | false | null | t3_1i7823a | /r/LocalLLaMA/comments/1i7823a/stargate_any_reason_to_be_optimistic_here_for/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'SYq4GjrSRFtv6nNi8-46v83gBQCgWS-yX2tjnQIvhLk', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/fpQ8njT_DRG46SuSohSLfEZQ1Lh_u_zuWHNrje9mZWs.jpg?width=108&crop=smart&auto=webp&s=56539b3fe6ab692fb77f34f81b934af2eef6b770', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/fpQ8njT_DRG46SuSohSLfEZQ1Lh_u_zuWHNrje9mZWs.jpg?auto=webp&s=5d4eeeb07362522c8536fc845025c40577ea68df', 'width': 200}, 'variants': {}}]} |
Stargate - Any reason to be optimistic here..? Or is it just a boost in the arm for ClosedAI? | 1 | [removed] | 2025-01-22T10:23:56 | https://www.reddit.com/r/LocalLLaMA/comments/1i783za/stargate_any_reason_to_be_optimistic_here_or_is/ | mark-lord | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i783za | false | null | t3_1i783za | /r/LocalLLaMA/comments/1i783za/stargate_any_reason_to_be_optimistic_here_or_is/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'SYq4GjrSRFtv6nNi8-46v83gBQCgWS-yX2tjnQIvhLk', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/fpQ8njT_DRG46SuSohSLfEZQ1Lh_u_zuWHNrje9mZWs.jpg?width=108&crop=smart&auto=webp&s=56539b3fe6ab692fb77f34f81b934af2eef6b770', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/fpQ8njT_DRG46SuSohSLfEZQ1Lh_u_zuWHNrje9mZWs.jpg?auto=webp&s=5d4eeeb07362522c8536fc845025c40577ea68df', 'width': 200}, 'variants': {}}]} |
Stargate - Any reason to be optimistic here..? | 1 | [removed] | 2025-01-22T10:24:27 | https://www.reddit.com/r/LocalLLaMA/comments/1i7847m/stargate_any_reason_to_be_optimistic_here/ | mark-lord | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7847m | false | null | t3_1i7847m | /r/LocalLLaMA/comments/1i7847m/stargate_any_reason_to_be_optimistic_here/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'SYq4GjrSRFtv6nNi8-46v83gBQCgWS-yX2tjnQIvhLk', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/fpQ8njT_DRG46SuSohSLfEZQ1Lh_u_zuWHNrje9mZWs.jpg?width=108&crop=smart&auto=webp&s=56539b3fe6ab692fb77f34f81b934af2eef6b770', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/fpQ8njT_DRG46SuSohSLfEZQ1Lh_u_zuWHNrje9mZWs.jpg?auto=webp&s=5d4eeeb07362522c8536fc845025c40577ea68df', 'width': 200}, 'variants': {}}]} |
Stargate - Any reason to be optimistic here for local models..? Or is it just a boost in the arm for ClosedAI? | 9 | **TL;DR -**
**Is closed vs open source zero sum?**
**What if OAI succeeds at regulatory capture, what then?**
**Will some of the $500bn go to orgs that release open weights e.g. Meta?**
**---**
Hey everyone!
I imagine many here will probably have the same initial reaction I did to the news - which was *uh oh.* But on the flip-side of the coin I've actually seen mostly positive reactions on my {bird-app} timeline. So I thought I'd do a vibe check / start a discussion here specifically, since the ripples made by this kind of project will for sure trickle down into the local scene. So:
**Is Stargate going to be good for local models?** | 2025-01-22T10:27:12 | https://www.reddit.com/r/LocalLLaMA/comments/1i785kc/stargate_any_reason_to_be_optimistic_here_for/ | mark-lord | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i785kc | false | null | t3_1i785kc | /r/LocalLLaMA/comments/1i785kc/stargate_any_reason_to_be_optimistic_here_for/ | false | false | self | 9 | null |
How easy is Fine-Tuning ? | 1 | [removed] | 2025-01-22T10:27:59 | https://www.reddit.com/r/LocalLLaMA/comments/1i785xn/how_easy_is_finetuning/ | Chance_Break6628 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i785xn | false | null | t3_1i785xn | /r/LocalLLaMA/comments/1i785xn/how_easy_is_finetuning/ | false | false | self | 1 | null |
DeepSeek using intuition vs logic | 10 | 2025-01-22T10:39:38 | https://www.reddit.com/gallery/1i78bp3 | DeviantPlayeer | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1i78bp3 | false | null | t3_1i78bp3 | /r/LocalLLaMA/comments/1i78bp3/deepseek_using_intuition_vs_logic/ | false | false | 10 | null |
||
R1 Distill + Rag | 6 | Did anybody already try this? R1 14b gave me always almost working complicated and nuanced python code but sometimes with wrong function names. When I input the api doc from packages it always stunned me how good it is. So before I work on a rag solution, did somebody already try it? I think a beautiful solution would be agentic, that an agent retrieves all necessary package documentation and sets it as the additional context | 2025-01-22T10:44:55 | https://www.reddit.com/r/LocalLLaMA/comments/1i78ebv/r1_distill_rag/ | Flashy_Management962 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i78ebv | false | null | t3_1i78ebv | /r/LocalLLaMA/comments/1i78ebv/r1_distill_rag/ | false | false | self | 6 | null |
Venting | 1 | [removed] | 2025-01-22T10:47:19 | https://www.reddit.com/r/LocalLLaMA/comments/1i78fir/venting/ | Upset_Acanthaceae_18 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i78fir | false | null | t3_1i78fir | /r/LocalLLaMA/comments/1i78fir/venting/ | false | false | self | 1 | null |
R2 when (your opinion) | 1 | [removed] | 2025-01-22T11:05:31 | https://www.reddit.com/r/LocalLLaMA/comments/1i78p3q/r2_when_your_opinion/ | Optimal_Hamster5789 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i78p3q | false | null | t3_1i78p3q | /r/LocalLLaMA/comments/1i78p3q/r2_when_your_opinion/ | false | false | self | 1 | null |
R2 will be disaster to ClosedAI | 1 | [removed] | 2025-01-22T11:08:44 | https://www.reddit.com/r/LocalLLaMA/comments/1i78qsl/r2_will_be_disaster_to_closedai/ | Optimal_Hamster5789 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i78qsl | false | null | t3_1i78qsl | /r/LocalLLaMA/comments/1i78qsl/r2_will_be_disaster_to_closedai/ | false | false | self | 1 | null |
Deepseek R1 GRPO code open sourced 🤯 | 1 | 2025-01-22T11:09:36 | https://www.reddit.com/r/LocalLLaMA/comments/1i78r8p/deepseek_r1_grpo_code_open_sourced/ | eliebakk | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i78r8p | false | null | t3_1i78r8p | /r/LocalLLaMA/comments/1i78r8p/deepseek_r1_grpo_code_open_sourced/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'TaCioUI7kR6Y8_CK8HdPTZc7J1hJOK1UbB54bvqtjgE', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/geCTeSIVMwNFsc6dmRyWBPp_gEBb1U86STRBWfXJy6w.jpg?width=108&crop=smart&auto=webp&s=b110b8a840fd1d9ec305b02caa89b6df67604d1a', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/geCTeSIVMwNFsc6dmRyWBPp_gEBb1U86STRBWfXJy6w.jpg?width=216&crop=smart&auto=webp&s=1417397f29b70de653d939a4d9f0e498f7859d24', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/geCTeSIVMwNFsc6dmRyWBPp_gEBb1U86STRBWfXJy6w.jpg?width=320&crop=smart&auto=webp&s=a4648d1d590c2b42c7312aeb31d9ae92e10a6d85', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/geCTeSIVMwNFsc6dmRyWBPp_gEBb1U86STRBWfXJy6w.jpg?width=640&crop=smart&auto=webp&s=2318e8b2a495fb8730a49369b4564294ffa6c271', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/geCTeSIVMwNFsc6dmRyWBPp_gEBb1U86STRBWfXJy6w.jpg?width=960&crop=smart&auto=webp&s=5403525e2285d6ee8c82228d33cef0f85e315ee0', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/geCTeSIVMwNFsc6dmRyWBPp_gEBb1U86STRBWfXJy6w.jpg?width=1080&crop=smart&auto=webp&s=5d5e682eebace4272424fd7fda95c743e49f9fe8', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/geCTeSIVMwNFsc6dmRyWBPp_gEBb1U86STRBWfXJy6w.jpg?auto=webp&s=cd82763f753c2aee0e5e320b93e7076c862f438e', 'width': 1200}, 'variants': {}}]} |
||
Deepseek R1 GRPO code open sourced 🤯 | 362 | 2025-01-22T11:11:56 | eliebakk | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i78sfs | false | null | t3_1i78sfs | /r/LocalLLaMA/comments/1i78sfs/deepseek_r1_grpo_code_open_sourced/ | false | false | 362 | {'enabled': True, 'images': [{'id': 'clhNbLsBecIiwYwQ18HfLTUbMHr94iaEx72pXDMvCVI', 'resolutions': [{'height': 57, 'url': 'https://preview.redd.it/ryfnofs83jee1.png?width=108&crop=smart&auto=webp&s=008f449a57b38c98751f7c236f56a13112e41bf5', 'width': 108}, {'height': 114, 'url': 'https://preview.redd.it/ryfnofs83jee1.png?width=216&crop=smart&auto=webp&s=ca6aa78423f1918c934fd296107d2a058782515c', 'width': 216}, {'height': 168, 'url': 'https://preview.redd.it/ryfnofs83jee1.png?width=320&crop=smart&auto=webp&s=30094fb20beb43002b7b0d9d0457aaac820f0436', 'width': 320}, {'height': 337, 'url': 'https://preview.redd.it/ryfnofs83jee1.png?width=640&crop=smart&auto=webp&s=40e8bb4b0b8a7dd1b82c628a825c88559a17aff0', 'width': 640}, {'height': 506, 'url': 'https://preview.redd.it/ryfnofs83jee1.png?width=960&crop=smart&auto=webp&s=0221000e72520d27e419bd5bf877eb750de9edba', 'width': 960}, {'height': 570, 'url': 'https://preview.redd.it/ryfnofs83jee1.png?width=1080&crop=smart&auto=webp&s=44d033510c37cfd0e4a108f9e4ca14616a7f994b', 'width': 1080}], 'source': {'height': 1861, 'url': 'https://preview.redd.it/ryfnofs83jee1.png?auto=webp&s=2466e78558db98eb350061d44614d65502897c1b', 'width': 3525}, 'variants': {}}]} |
|||
How I Used GPT-O1 Pro to Discover My Autoimmune Disease (After Spending $100k and Visiting 30+ Hospitals with No Success) | 499 | **TLDR:**
* **Suffered from various health issues for 5 years, visited 30+ hospitals with no answers**
* **Finally diagnosed with axial spondyloarthritis through genetic testing**
* **Built a personalized health analysis system using GPT-O1 Pro, which actually suggested this condition earlier**
I'm a guy in my mid-30s who started having weird health issues about 5 years ago. Nothing major, but lots of annoying symptoms - getting injured easily during workouts, slow recovery, random fatigue, and sometimes the pain was so bad I could barely walk.
At first, I went to different doctors for each symptom. Tried everything - MRIs, chiropractic care, meds, steroids - nothing helped. I followed every doctor's advice perfectly. Started getting into longevity medicine thinking it might be early aging. Changed my diet, exercise routine, sleep schedule - still no improvement. The cause remained a mystery.
Recently, after a month-long toe injury wouldn't heal, I ended up seeing a rheumatologist. They did genetic testing and boom - diagnosed with axial spondyloarthritis. This was the answer I'd been searching for over 5 years.
Here's the crazy part - I fed all my previous medical records and symptoms into GPT-O1 pro before the diagnosis, and it actually listed this condition as the top possibility!
This got me thinking - why didn't any doctor catch this earlier? Well, it's a rare condition, and autoimmune diseases affect the whole body. Joint pain isn't just joint pain, dry eyes aren't just eye problems. The usual medical workflow isn't set up to look at everything together.
So I had an idea: What if we created an open-source system that could analyze someone's complete medical history, including family history (which was a huge clue in my case), and create personalized health plans? It wouldn't replace doctors but could help both patients and medical professionals spot patterns.
Building my personal system was challenging:
1. Every hospital uses different formats and units for test results. Had to create a GPT workflow to standardize everything.
2. RAG wasn't enough - needed a large context window to analyze everything at once for the best results.
3. Finding reliable medical sources was tough. Combined official guidelines with recent papers and trusted YouTube content.
4. GPT-O1 pro was best at root cause analysis, Google Note LLM worked great for citations, and Examine excelled at suggesting actions.
In the end, I built a system using Google Sheets to view my data and interact with trusted medical sources. It's been incredibly helpful in managing my condition and understanding my health better.
| 2025-01-22T11:12:33 | https://www.reddit.com/r/LocalLLaMA/comments/1i78src/how_i_used_gpto1_pro_to_discover_my_autoimmune/ | Dry_Steak30 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i78src | false | null | t3_1i78src | /r/LocalLLaMA/comments/1i78src/how_i_used_gpto1_pro_to_discover_my_autoimmune/ | false | false | self | 499 | null |
Deepseek-r1 going rogue with Ollama | 1 | This just happened when I tested the new version of deepseek-r1 8b on Ollama with some basic query about its own CoT process as well as a random question about movie release date to see how the CoT would process something involving fetching a specific date. I was simultaneously running some pip installs on VS Code when it seems Ollama suddenly picked up on it, causing deepseek to have nonstop CoT. Was quite freaky to be honest. Anyone seen something similar? It just went on and on, the log of one terminal was feeding input to Ollama... | 2025-01-22T11:37:33 | https://www.reddit.com/r/LocalLLaMA/comments/1i795wf/deepseekr1_going_rogue_with_ollama/ | cybercyrus | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i795wf | false | null | t3_1i795wf | /r/LocalLLaMA/comments/1i795wf/deepseekr1_going_rogue_with_ollama/ | false | false | self | 1 | null |
Best offline model for commercial use | 1 | Hi, I'm currently working to an AI agent which will use an LLM.
What is/are the best LLM to be used locally?
I need something which can be used for commercial purposes, can be fine tuned and run locally.
Also what kind of GPU is needed to only make inferences or also to fine-tune that LLMs?
Thank you all in advance. | 2025-01-22T11:49:11 | https://www.reddit.com/r/LocalLLaMA/comments/1i79cav/best_offline_model_for_commercial_use/ | Giux99 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i79cav | false | null | t3_1i79cav | /r/LocalLLaMA/comments/1i79cav/best_offline_model_for_commercial_use/ | false | false | self | 1 | null |
A question about DeepSeek R1 pricing and in general reasoning based LLM pricing | 1 | [removed] | 2025-01-22T12:34:57 | https://www.reddit.com/r/LocalLLaMA/comments/1i7a428/a_question_about_deepseek_r1_pricing_and_in/ | liveforfiction | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7a428 | false | null | t3_1i7a428 | /r/LocalLLaMA/comments/1i7a428/a_question_about_deepseek_r1_pricing_and_in/ | false | false | self | 1 | null |
Repurposing ETH Mining Rigs for AI Workloads | 1 | [removed] | 2025-01-22T12:42:01 | [deleted] | 1970-01-01T00:00:00 | 0 | {} | 1i7a8h0 | false | null | t3_1i7a8h0 | /r/LocalLLaMA/comments/1i7a8h0/repurposing_eth_mining_rigs_for_ai_workloads/ | false | false | default | 1 | null |
||
Repurposing ETH Mining Rigs for AI Workloads | 1 | [removed] | 2025-01-22T12:45:07 | https://www.reddit.com/r/LocalLLaMA/comments/1i7aaed/repurposing_eth_mining_rigs_for_ai_workloads/ | Gioxxy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7aaed | false | null | t3_1i7aaed | /r/LocalLLaMA/comments/1i7aaed/repurposing_eth_mining_rigs_for_ai_workloads/ | false | false | self | 1 | null |
Can DeepSeek-R1 be used as a better frontend for Claude? | 1 | [removed] | 2025-01-22T12:46:45 | https://www.reddit.com/r/LocalLLaMA/comments/1i7abfo/can_deepseekr1_be_used_as_a_better_frontend_for/ | dT4YGAKxT6 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7abfo | false | null | t3_1i7abfo | /r/LocalLLaMA/comments/1i7abfo/can_deepseekr1_be_used_as_a_better_frontend_for/ | false | false | self | 1 | null |
Lightweight web scraping with ollama | 32 | Hi LocalLLaMA people,
It’s been almost 6 months since I’ve shared my lightweight library for easy web scraping. Today, I want to share some updates, starting with the fact that **Parsera now officially supports Ollama!**
Some notable changes include:
* **Cookies support** for accessing login-required content
* **Enhanced page rendering** to ensure no content is missed
* **Custom script execution**, allowing scraping of interactive elements
* **Page chunking**, to process pages that exceed the model’s context limits
* Plus many other minor enhancements and fixes
Check it out on GitHub: [https://github.com/raznem/parsera](https://github.com/raznem/parsera)
P. S. We are now using Parsera as the foundation for our upcoming Scraping Agent, so hope I’ll share something exciting soon. | 2025-01-22T12:49:48 | https://www.reddit.com/r/LocalLLaMA/comments/1i7adec/lightweight_web_scraping_with_ollama/ | Financial-Article-12 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7adec | false | null | t3_1i7adec | /r/LocalLLaMA/comments/1i7adec/lightweight_web_scraping_with_ollama/ | false | false | self | 32 | {'enabled': False, 'images': [{'id': 'XiHC3l1zAwCmGAPyzE3JfRwMM5v6mYJtQnXJi7og_o4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/36rF9YETkivuXVSIbLTODTa_A-EbXoZPq5Xu4i60tvI.jpg?width=108&crop=smart&auto=webp&s=233ec7cfa27470418f75f82253b929ff348aa5c7', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/36rF9YETkivuXVSIbLTODTa_A-EbXoZPq5Xu4i60tvI.jpg?width=216&crop=smart&auto=webp&s=49045dc7d0fe923666d396b9e0efaf29a2ed2456', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/36rF9YETkivuXVSIbLTODTa_A-EbXoZPq5Xu4i60tvI.jpg?width=320&crop=smart&auto=webp&s=c8530888e2412082113e83ba3d5362e0a095a99f', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/36rF9YETkivuXVSIbLTODTa_A-EbXoZPq5Xu4i60tvI.jpg?width=640&crop=smart&auto=webp&s=54ff04bfd8df4a7da2bf5985feb426a2f64f0f5c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/36rF9YETkivuXVSIbLTODTa_A-EbXoZPq5Xu4i60tvI.jpg?width=960&crop=smart&auto=webp&s=7b59cfc3b3f5e305ff2799bc7038ca50cdb5ec2f', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/36rF9YETkivuXVSIbLTODTa_A-EbXoZPq5Xu4i60tvI.jpg?width=1080&crop=smart&auto=webp&s=5c20758c6dc5249ec45abe9073bd2f5389015d0b', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/36rF9YETkivuXVSIbLTODTa_A-EbXoZPq5Xu4i60tvI.jpg?auto=webp&s=a5d6f0c66cc602f31245705aec018366ef25aa3c', 'width': 1200}, 'variants': {}}]} |
I am compelled to create this post (Extreme rant: very sensitive!) | 1 | [removed] | 2025-01-22T13:04:21 | https://www.reddit.com/r/LocalLLaMA/comments/1i7an1p/i_am_compelled_to_create_this_post_extreme_rant/ | BLsaw | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7an1p | false | null | t3_1i7an1p | /r/LocalLLaMA/comments/1i7an1p/i_am_compelled_to_create_this_post_extreme_rant/ | false | false | nsfw | 1 | null |
Editing <thinking> process is a super effective way to guide model output | 30 | R1 is my first real foray into reasoning models. One thing I have gotten in the habit of with Claude is to edit my prior prompt to guide model output, rather than engaging in a long back and forth. This is even more powerful with reasoning models, as the thinking tags act as a scaffold for model output. You can edit, or simply add on, new thinking content in the reasoning section and re-run the completion to get what you want. In my experience, it is a bit more efficient than modifying the prompt directly, as it's already interpreted the prompt and so the thinking tags are closer to the output than the prompt is (so less of a chance to misinterpret the user instructions, and more of an opportunity to subtly guide output).
Here is an example. I'm a scientist, so my default test of the model is to have it output a classic model in evolutionary biology, the Price Equation applied to multilevel selection. The details don't really matter here, but in the revision, I simply added a thinking tag telling it to output a numerical example. This is a coarse example that definitely could have been accomplished in the prompt, but I figured for this example something coarse is probably easier to understand by those unfamiliar with the subject than something subtle. Still, you get the idea!
For this test I am using R1-70b distill, q8, Mac Studio in LM Studio.
**Version 1.**
**User**
Use the Price equation to partition selection acting on individual bees in a colony from that acting on the colony as a whole.
**Assistant**
**<think>**First, I remember that the Price equation is used in evolutionary biology to partition selection into different components. It's about HOW selection acts on traits, whether at the individual level or group level.
So, applying it here: we have a honeybee colony with many workers. Each bee has some trait—maybe foraging behavior—that varies among them. The success of the colony depends on how good the bees are at this trait.
The Price equation basically breaks down change in a trait into two parts: selection and transmission. Wait, no—I think it's more about genetic variation within groups and between groups. Oh right, the equation is Δz = V\_S + VT.
Wait, actually, the Price equation formula is typically written as:
Δz̄ = (Cov(w, z) / w̄ )
Where w is fitness, z is the trait. But when applied hierarchically, it can be split into within-group and between-group components, right?
So in this case, individual selection would act on the variation among bees within a colony, affecting their relative contributions (maybe how many offspring each contributes). Colony-level selection would be about how well each colony does compared to others based on average traits.
For honeybees, maybe it's about how efficient foragers are. Some colonies have more efficient workers on average, so those colonies survive or propagate better. But within each colony, bees with better traits might contribute more via reproduction, but wait, in bees, queens are the ones who reproduce, not workers. Oh, that complicates things.
Wait, so worker bees don't have direct reproduction since they're sterile. So individual selection on them is...Hmm. Maybe this is a way to model group-level adaptation through individual contributions, even if those individuals aren't reproducing themselves. Or perhaps the Price equation can still be applied to see how colony-level fitness aggregates from individual behaviors.
So structure-wise: Each colony has bees with trait z\_i. The average of z\_i in a colony is Z\_j. Colony j's fitness W\_j depends on Z\_j.
The overall change in trait would come from two sources: selection within colonies (how each bee's trait affects the colony's success) and selection between colonies based on their averages.
Putting it into equations, I need to show the Price equation for both individual and group levels.
But wait, since worker bees don't reproduce directly, what constitutes 'fitness' w for them? Maybe it's about how much they contribute to the colony's fitness. So each bee contributes to W\_j based on z\_i.
So in that case, within-colony selection is based on individual contributions to Wj, and between-colony selection is based on colonies with higher average Zj having higher fitness.
Yeah, so I think that's how it breaks down: part of the equation captures variation among bees in how they help their colony (within-colony), and another part captures how colonies' averages affect their success (between-colonies). **</think>**
To apply the Price Equation to partition selection acting on individual bees within a hive versus selection at the colony level, we consider the following model:
Let ( z\_i ) represent an altruistic trait of bee ( i ) in colony ( j ). The fitness ( w\_{i} ) contributes to mortality/survival rates depending on this trait. Colony reproductive success is proportional to ( w\_j = average(w\_i for all bees in colony j) ).
The Price Equation quantifies the change in mean trait from one generation to the next as:
\[ \\Delta z̄ = \\frac{\\text{Cov}(z, w)}{\\bar{w}} + (\\bar{w}\_j - \\bar{w}) \]
Here, the first term represents individual-level selection: covariation between bee trait ( z\_i ) and their contribution to colony success. The second term is colony-level selection, where colonies with higher average ( w\_j ) contribute disproportionately more to new colonies.
Thus, selection on bees has two components:
1. **Within-colony**: The effect of each bee's individual traits on its own survival/efficiency.
2. **Between-colonies**: The fitness difference between colonies due to their collective trait averages.
This partition highlights how both levels contribute to evolutionary change in the population, even if worker bees themselves do not reproduce directly.
**Version 2**
I just added this tot he end of the thinking tags, deleted the output below the thinking tags, and then re-ran the completion in LM Studio:
**Once I have derived the general case, I will provide a fully worked numerical example, to make it clear how this is used. </think>**
To apply the Price equation to understand selection acting on individual bees versus the colony as a whole, we need to partition selective forces into within-group and between-group components. Let’s consider a population of honeybee colonies where each colony comprises multiple worker bees exhibiting some heritable trait ( z ) (e.g., foraging efficiency). The fitness of a colony ( j ), denoted ( W\_j ), depends on the average value of ( z ) across its workers, as well as potentially other factors such as environmental conditions and inter-colony competition.
**Step 1: Define Variables**
* Let ( Z\_j = \\frac{1}{N\_j} \\sum\_{i=1}\^{N\_j} z\_i ), where ( N\_j ) is the number of workers in colony ( j ).
* The aggregate fitness of colony ( j ) is some function of ( Z\_j ), i.e., ( W\_j = f(Z\_j) ).
**Step 2: Apply the Price Equation**
The standard form of the Price equation is:
\[ \\Delta z = \\frac{1}{\\overline{W}} \\sum\_{j=1}\^{M} (W\_j - \\overline{W}) Z\_j + \\frac{1}{\\overline{W}} \\sum\_{j=1}\^{M} W\_j \\left( z\_j - Z\_j \\right). \]
Here, ( z ) is the trait of interest, ( W ) is fitness, and a bar denotes an average over all colonies.
**Step 3: Interpretation**
The first term in the equation represents selection acting at the colony level (between-group selection). This measures how differences in the average trait (( Z\_j )) between colonies contribute to differences in colony fitness (( W\_j )). Colonies with higher values of ( Z\_j ) will have higher fitness, so this term quantifies multilevel selection at the colony level.
The second term represents within-group selection. This measures how much individual deviations from the group mean (( z\_j - Z\_j )) are associated with differences in individual contributions to colony success. If certain workers contribute more to colony-level traits (e.g., foraging efficiency), this difference contributes to anisman evolutionary outcome at this level.
**Step 4: Connection Back to Honeybees**
For honeybees, the within-colony term can be thought of as selection acting on worker behaviors that enhance colony success. For example, more efficient foragers contribute disproportionately to colony resources and survival. The between-colony term reflects differences in overall colony performance based on the average traits of their workers.
By partitioning these effects using the Price equation, we can quantify the relative strength of individual versus group selection in shaping evolutionary outcomes in social bees.
**Step 5: Numerical Example**
Assume we have three colonies with different mean foraging efficiencies (( Z\_1 = 3 ), ( Z\_2 = 4 ), ( Z\_3 = 5 )) on a scale where higher values represent better efficiency. Let the fitness of each colony be ( W\_1 = 10 ), ( W\_2 = 20 ), and ( W\_3 = 30 ). Suppose the mean fitness across all colonies is ( \\overline{W} = (10 + 20 + 30)/3 = \\frac{60}{3} = 20 ).
Compute each term:
1. **Between-group selection:**
\[ \\frac{1}{\\overline{W}} \\sum\_{j=1}\^{M} (W\_j - \\overline{W}) Z\_j = \\frac{1}{20} \[(10-20) \\cdot 3 + (20-20) \\cdot 4 + (30-20) \\cdot 5\] = \\frac{1}{20} \[(-10)\\cdot3 + 0\\cdot4 + 10\\cdot5\] = \\frac{1}{20} (-30 + 50) = \\frac{20}{20} = 1. \]
1. **Within-group selection:**
Assume within-colony deviations (( z\_j - Z\_j )) are uncorrelated with colony fitness for simplicity, so this term is negligible (or zero).
Thus, the overall change in ( z ) comes entirely from between-group selection:( \\Delta z = 1 + 0 = 1 ). | 2025-01-22T13:06:20 | https://www.reddit.com/r/LocalLLaMA/comments/1i7aoet/editing_thinking_process_is_a_super_effective_way/ | Berberis | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7aoet | false | null | t3_1i7aoet | /r/LocalLLaMA/comments/1i7aoet/editing_thinking_process_is_a_super_effective_way/ | false | false | self | 30 | null |
Memory bandwidth of Nvidia RTX Laptop graphics compared | 52 | 2025-01-22T13:09:26 | Balance- | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i7aqhg | false | null | t3_1i7aqhg | /r/LocalLLaMA/comments/1i7aqhg/memory_bandwidth_of_nvidia_rtx_laptop_graphics/ | false | false | 52 | {'enabled': True, 'images': [{'id': 'CByCxQC2bBTqJ5yzvR9DeSLj68FYO_8QU7xofAuqfHI', 'resolutions': [{'height': 65, 'url': 'https://preview.redd.it/3ec8i695ojee1.png?width=108&crop=smart&auto=webp&s=dfbd109c576d2e6d4c142a06c0e579d82d962c9a', 'width': 108}, {'height': 131, 'url': 'https://preview.redd.it/3ec8i695ojee1.png?width=216&crop=smart&auto=webp&s=a7e84c52ec38eb433870bd9021a4dfea80bb4943', 'width': 216}, {'height': 194, 'url': 'https://preview.redd.it/3ec8i695ojee1.png?width=320&crop=smart&auto=webp&s=186988e87ba70b0a4287e18317ffcef2e456ffba', 'width': 320}, {'height': 388, 'url': 'https://preview.redd.it/3ec8i695ojee1.png?width=640&crop=smart&auto=webp&s=3fbd835d18c9f6df088cfd0931409590eaf41d47', 'width': 640}, {'height': 582, 'url': 'https://preview.redd.it/3ec8i695ojee1.png?width=960&crop=smart&auto=webp&s=fe471d32a654f1751865a1abdf6126d2ba8dd13d', 'width': 960}, {'height': 655, 'url': 'https://preview.redd.it/3ec8i695ojee1.png?width=1080&crop=smart&auto=webp&s=157d8c8282544bad5da65a88d9bb3409354b3a8e', 'width': 1080}], 'source': {'height': 1342, 'url': 'https://preview.redd.it/3ec8i695ojee1.png?auto=webp&s=99a83b4e0154d6ce7561bfcb30631aeef9501c69', 'width': 2212}, 'variants': {}}]} |
|||
About DeepSeek R1 Zero | 1 | [removed] | 2025-01-22T13:17:10 | https://www.reddit.com/r/LocalLLaMA/comments/1i7avut/about_deepseek_r1_zero/ | Berion-Reviador | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7avut | false | null | t3_1i7avut | /r/LocalLLaMA/comments/1i7avut/about_deepseek_r1_zero/ | false | false | self | 1 | null |
Suggestion needed: future proofing set up | 1 | [removed] | 2025-01-22T13:22:10 | https://www.reddit.com/r/LocalLLaMA/comments/1i7azcq/suggestion_needed_future_proofing_set_up/ | kpolany | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7azcq | false | null | t3_1i7azcq | /r/LocalLLaMA/comments/1i7azcq/suggestion_needed_future_proofing_set_up/ | false | false | self | 1 | null |
I did a quick test of MacBook M4 Max 128 GB token/second throughput across a few popular local LLMs (in the MLX format) | 87 | I'm sharing this in case you were wondering what kind of throughput you might expect to get on a machine like this. E.g. if you are considering whether it's worth buying or not (as for me, I have no regrets, I'm loving this beast). Same single query, set the context to 8K, tok/sec numbers are reported below, as measured by LMStudio:
LLaMA 3.2 3B 4bit -- 181
LLaMA 3 8B 8bit -- 55
LLaMA 3.3 70B 4bit -- 11.8
Mistral Large 123B 4bit -- 6.6
Mistral Nemo 12B 4bit -- 63
Mistral Nemo 12B 8bit -- 36
Mistral Small 22B 4bit -- 34.5
Mistral Small 22B 8bit -- 19.6
Qwen2.5 32B 4bit -- 24
Qwen2.5 32B 8bit -- 13.5
Qwen2.5 72B 4bit -- 10.9
Qwen2.5 72B 8bit -- 6.2
WizardLM-2 8x22B 4bit -- 19.4!!
For comparison, here are some numbers obtained in the same setting on my other MacBook, M1 Pro with 32 GB:
Mistral Nemo 12B 4bit -- 22.8
Mistral Small 22B 4bit -- 12.9
Qwen2.5 32B 4bit -- 8.8
Hope it's interesting / useful. | 2025-01-22T13:28:52 | https://www.reddit.com/r/LocalLLaMA/comments/1i7b3r1/i_did_a_quick_test_of_macbook_m4_max_128_gb/ | Pure_Refrigerator988 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7b3r1 | false | null | t3_1i7b3r1 | /r/LocalLLaMA/comments/1i7b3r1/i_did_a_quick_test_of_macbook_m4_max_128_gb/ | false | false | self | 87 | null |
DeepSeek in k8s | 3 | We have 6 H100 servers, as you know each one have 8 (80GB)GPUs, The biggest model we have is llama 405B which takes 8 GPU's which fits one server. now deep seek needs more than 8 GPU's does anybody know a way to run the model on multiple H100 servers?
ps: our stack fully k8s | 2025-01-22T13:34:15 | https://www.reddit.com/r/LocalLLaMA/comments/1i7b7hp/deepseek_in_k8s/ | No-Emphasis6569 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7b7hp | false | null | t3_1i7b7hp | /r/LocalLLaMA/comments/1i7b7hp/deepseek_in_k8s/ | false | false | self | 3 | null |
Serverless Endpoint /w deepseek-ai/DeepSeek-R1 on Runpod? | 1 | [removed] | 2025-01-22T13:38:51 | https://www.reddit.com/r/LocalLLaMA/comments/1i7bal5/serverless_endpoint_w_deepseekaideepseekr1_on/ | LLattus | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7bal5 | false | null | t3_1i7bal5 | /r/LocalLLaMA/comments/1i7bal5/serverless_endpoint_w_deepseekaideepseekr1_on/ | false | false | self | 1 | null |
Help with model for information extraction. | 2 | Hello folks!
I'm new to LLM usage aside from chat/assistents things.
I'm trying to create an automation which would receive and unstructured file and it would output the information that I need in a structured format (like JSON).
I tried Phi4 and Qwen2:7b models, but they don't seems to output what I ask, they just give me a summary.
I good pretty good results using Gemini-2.0-flash-exp, but the output is only 8k, which is unsuitable for processing files, which in average, have 200k tokens.
So I was wondering if there is a model which could do this task and also has a bigger output size, like 128k or more.
Thanks! | 2025-01-22T13:49:25 | https://www.reddit.com/r/LocalLLaMA/comments/1i7bi1w/help_with_model_for_information_extraction/ | afonsolage | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7bi1w | false | null | t3_1i7bi1w | /r/LocalLLaMA/comments/1i7bi1w/help_with_model_for_information_extraction/ | false | false | self | 2 | null |
New AI IDE, Trae | 1 | 2025-01-22T13:50:06 | https://v.redd.it/gq6tnnwhvjee1 | ozgrozer | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1i7bij4 | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/gq6tnnwhvjee1/DASHPlaylist.mpd?a=1740145822%2CZWI2NzliODY0MDYyMTlmOWFiODNkZDY3NTA1MmJmYjZlZWVlZWRhYWVmMWYxOGE0MmYyMjhjNjhhYWI2N2VhNg%3D%3D&v=1&f=sd', 'duration': 25, 'fallback_url': 'https://v.redd.it/gq6tnnwhvjee1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/gq6tnnwhvjee1/HLSPlaylist.m3u8?a=1740145822%2CNWI1NDE0YTE1MDM3NTk0NTY4NmE5NTM2YTg1OWU4ZWFlMjJhZTg0MGIyMWJhMTRjOTE3MzA1OGQ5Y2ZlYzY0MQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/gq6tnnwhvjee1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1600}} | t3_1i7bij4 | /r/LocalLLaMA/comments/1i7bij4/new_ai_ide_trae/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-', 'resolutions': [{'height': 72, 'url': 'https://external-preview.redd.it/aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-.png?width=108&crop=smart&format=pjpg&auto=webp&s=ff6e02136c014b5a7c1c7c6728fbaa0523eea37f', 'width': 108}, {'height': 145, 'url': 'https://external-preview.redd.it/aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-.png?width=216&crop=smart&format=pjpg&auto=webp&s=b662687d82ec19d8cf0b2bf1b43f92de6d067d57', 'width': 216}, {'height': 216, 'url': 'https://external-preview.redd.it/aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-.png?width=320&crop=smart&format=pjpg&auto=webp&s=38e190ac2eec74bd34d4945d4cfab7dd5b9d5110', 'width': 320}, {'height': 432, 'url': 'https://external-preview.redd.it/aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-.png?width=640&crop=smart&format=pjpg&auto=webp&s=5591c1932d2fd27c7891e1e0cfa107223f2f0309', 'width': 640}, {'height': 648, 'url': 'https://external-preview.redd.it/aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-.png?width=960&crop=smart&format=pjpg&auto=webp&s=43694bdff15d5f9adb2629c39bbea8133e274470', 'width': 960}, {'height': 729, 'url': 'https://external-preview.redd.it/aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-.png?width=1080&crop=smart&format=pjpg&auto=webp&s=665205782edaf3f48a64c530ad1557b7a03356c5', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/aG9pcHdud2h2amVlMc2joKsfluBg59-vRd-_kihMffjdNW8R1p-eEeOoKZY-.png?format=pjpg&auto=webp&s=4c7dc93302808e516f7e13bebe6495684269b726', 'width': 1600}, 'variants': {}}]} |
||
Any way to issue permanent commands? | 1 | [removed] | 2025-01-22T14:02:38 | https://www.reddit.com/r/LocalLLaMA/comments/1i7brny/any_way_to_issue_permanent_commands/ | foulplayjamm | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7brny | false | null | t3_1i7brny | /r/LocalLLaMA/comments/1i7brny/any_way_to_issue_permanent_commands/ | false | false | self | 1 | null |
Help me justify upgrading 64gb M1 Max MBP to 128gb M4? | 4 | I have a maxed out 64gb M1 Max MBP with 8TB of storage space
I always wanted more RAM, and in 2023 and early 2024 it was to run larger models
but now it seems like
A) inference in the cloud has gotten so cheap, plummeting 80-90% from the hourly cost in early 2023
B) smaller models have gotten so much better, for my super private offline conversations
C) resale value of the M1 Max is not nearly close to the $7k I paid for it, and the $7k I would need for a maxed out M4
D) the speed of RAM, processing, and other use cases don't seem thaaaat much of an improvement. like maybe $1,000 more but not $5,000 more
E) the extra large models wouldn't run on 128gb RAM either
so I do like new things, and this machine, the M1, will only get older. so I'm curious about other ways to justify it. like, what am I missing | 2025-01-22T14:05:58 | https://www.reddit.com/r/LocalLLaMA/comments/1i7bu7c/help_me_justify_upgrading_64gb_m1_max_mbp_to/ | thetaFAANG | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7bu7c | false | null | t3_1i7bu7c | /r/LocalLLaMA/comments/1i7bu7c/help_me_justify_upgrading_64gb_m1_max_mbp_to/ | false | false | self | 4 | null |
Mass email communication management using LLM | 1 | [removed] | 2025-01-22T14:13:18 | https://www.reddit.com/r/LocalLLaMA/comments/1i7bzmm/mass_email_communication_management_using_llm/ | No-Temporary9401 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7bzmm | false | null | t3_1i7bzmm | /r/LocalLLaMA/comments/1i7bzmm/mass_email_communication_management_using_llm/ | false | false | self | 1 | null |
Mass email storage, management and visualization | 1 | [removed] | 2025-01-22T14:16:52 | https://www.reddit.com/r/LocalLLaMA/comments/1i7c2c5/mass_email_storage_management_and_visualization/ | nabroski18 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7c2c5 | false | null | t3_1i7c2c5 | /r/LocalLLaMA/comments/1i7c2c5/mass_email_storage_management_and_visualization/ | false | false | self | 1 | null |
AMA with llama.cpp quantization god Iwan Kawrakow at FOSDEM | 1 | [removed] | 2025-01-22T14:27:59 | https://www.reddit.com/r/LocalLLaMA/comments/1i7cawz/ama_with_llamacpp_quantization_god_iwan_kawrakow/ | asf-rvs | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7cawz | false | null | t3_1i7cawz | /r/LocalLLaMA/comments/1i7cawz/ama_with_llamacpp_quantization_god_iwan_kawrakow/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'cgSmh06wEoky2xD_QDCRBAGsBvTpB_i0pA2iS6WSslI', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/P3lyfhnClpqOXoYr5hnW1nzsKvn-H05FYkcfYIDxqrU.jpg?width=108&crop=smart&auto=webp&s=e20724b837a14c3dc4539c2b1e94263a68911e9d', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/P3lyfhnClpqOXoYr5hnW1nzsKvn-H05FYkcfYIDxqrU.jpg?auto=webp&s=2f50f05544d563f3228af7801e140e17d8bbc2d4', 'width': 200}, 'variants': {}}]} |
Reasoning LLM vs standard LLM | 1 | [removed] | 2025-01-22T14:36:57 | https://www.reddit.com/r/LocalLLaMA/comments/1i7chs8/reasoning_llm_vs_standard_llm/ | FormerNegotiation428 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7chs8 | false | null | t3_1i7chs8 | /r/LocalLLaMA/comments/1i7chs8/reasoning_llm_vs_standard_llm/ | false | false | self | 1 | null |
AMD HX370 LLM performance | 6 | I just got an AMD HX 370-based MiniPC, and at this time (January 2025), it's not really suitable for serious LLM work. The NPU isn't supported even by AMD's ROCm, so it's basically useless.
CPU-based inference with ollama, with deepseek-r1:14b, results in 7.5 tok/s.
GPU-based inference with llama.cpp and the Vulkan API yields almost the same result, 7.8 tok/s (leaving CPU cores free to do other work).
The similarity of the results suggest that memory bandwidth is the probable bottleneck. I did these tests on a stock configuration with LPDDR5x 7500 MT/s, arranged in 4 channels of 8 GB, but the bus is 32-bit so it's like 128-bit total width. AIDA64 reports less than 90 GB/s memory read performance.
AMD calls it an "AI" chip, but - no it's not. At least not until drivers start supporting the NPU.
OTOH, by every other benchmark, it's blazing fast! | 2025-01-22T14:38:34 | https://www.reddit.com/r/LocalLLaMA/comments/1i7cj11/amd_hx370_llm_performance/ | ivoras | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7cj11 | false | null | t3_1i7cj11 | /r/LocalLLaMA/comments/1i7cj11/amd_hx370_llm_performance/ | false | false | self | 6 | null |
Running Phi-4 on NPU? | 2 | Hi guys,
I’m new to this and I would like to know if it’s possible to run the Phi-4 model from Microsoft on an NPU.
I installed this model on of my laptops having a 4090 RTX GPU and running fantastically.
I tried this model on my other laptop that doesn’t have a GPU but has an NPU and the NPU is not used at all by ollama. It’s only using the CPU and it’s too slow to be usable.
Any idea what I can do to my use of my NPU?
Thanks a lot in advance! | 2025-01-22T14:42:08 | https://www.reddit.com/r/LocalLLaMA/comments/1i7clss/running_phi4_on_npu/ | LocSta29 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7clss | false | null | t3_1i7clss | /r/LocalLLaMA/comments/1i7clss/running_phi4_on_npu/ | false | false | self | 2 | null |
Reasoning LLM vs Standard LLM | 5 | With the new "generation" of reasoning LLMs I'm starting asking myself: are always reasoning models the best fit?
I can understand the use of reasoning models for tasks like math or logic puzzles, but what about RAG (agentic or vanilla)? Are you comparing the precision of your RAG system with two comparable standard and reasoning LLMs?
DeepSeek V3 vs R1: which one do you use and why? Is V3 to be considered outdated? | 2025-01-22T14:49:19 | https://www.reddit.com/r/LocalLLaMA/comments/1i7cr85/reasoning_llm_vs_standard_llm/ | SK33LA | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1i7cr85 | false | null | t3_1i7cr85 | /r/LocalLLaMA/comments/1i7cr85/reasoning_llm_vs_standard_llm/ | false | false | self | 5 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.