Thomas Wolf, кофаундер HuggingFace, написал пост про пару вчерашних LLM анонсов. Нет смысла пересказывать это своими словами, оставлю as is: Taking a moment to detail yesterday's two new open-source AI model releases that I briefly mentioned in my long post. They are again pushing the state of the art! Small 3 by Mistral (Paris, France) is your top-level mid-sized model for fast-inference under Apache 2 licence. A 24B model made to run fast while keeping good performances (about GPT-4 level of performances in a model 70x smaller, generally about the perf/latency of today's GPT-4o-mini): Check it out at https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501 And read the blog post at https://mistral.ai/news/mistral-small-3/ Tülu 3 by AllenAI (Seattle, US) is your new very large open-source frontier model. With 405B parameters you won't run it everywhere for sure but according to the benchmarks of the AllenAI team it seems to outperform the base model being DeepSeek that everyone is talking about. Check it out at https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B And read the blog post at https://allenai.org/blog/tulu-3-405B By the way, both are the third versions released by these teams, and with the current base model from DeepSeek also being V3, seems like the whole open-source AI community is well aligned on versioning for some reason. Now the remaining part for both Small 3 and Tülu 3 will be to fine-tune these models following the DeepSeek recipe and turn them into the reasoning models we're starting to love. If you want to follow that endeavor, you can, for instance, head to our open-r1 GitHub repo, where we're reproducing the DeepSeek R1 reasoning recipe and extending it to many open-source models and domains. Exciting times to be alive! https://www.linkedin.com/posts/thom-wolf_taking-a-moment-to-detail-yesterdays-two-activity-7291061177445470208-XDIS Ждём открытых reasoning моделей вменяемого размера!