歡迎 Mixtral：Hugging Face 迎來頂尖的混合專家（MoE）開源模型
Original: Welcome Mixtral - a SOTA Mixture of Experts on Hugging Face
French AI startup Mistral AI officially released its highly anticipated open-source Mixture of Experts (MoE) model — Mixtral 8x7B. The…
Mistral AI 發表全新混合專家（MoE）模型 Mixtral 8x7B，總參數 46.7B 但每次 token 僅啟用 12.9B，速度極快。其在多項基準測試中超越 Llama 2 70B 及 GPT-3.5，並支援 32k 脈絡長度。Hugging Face 已於 Transformers、TGI 及推理終端全面整合，並採 Apache 2.0 開源授權。
French AI startup Mistral AI officially released its highly anticipated open-source Mixture of Experts (MoE) model — Mixtral 8x7B. The model caused a sensation in the open-source community, and Hugging Face immediately announced full support, making it easy for developers to deploy, fine-tune, and run inference on the model.
Free shows the 3-line summary; Pro unlocks the full deep summary (~300 words) so you never have to click through.
See Pro plans →Want the original English / full article?
Read on Hugging Face Blog →Summaries are AI-generated; the original article is authoritative.