[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-5b9047a1-4d24-444f-8f01-8f0d4c7d3767":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"5b9047a1-4d24-444f-8f01-8f0d4c7d3767","Mistral 3 全家桶正式开源：675B MoE 把「开放前沿」拉回牌桌，Apache 2.0 一口气打到边缘","法国 Mistral AI 今天正式发布 Mistral 3 家族，从云端旗舰到边缘小模型一口气全开源。旗舰 Mistral Large 3 采用稀疏 MoE 架构，总参数 675B、单 token 激活 41B，搭配 2.5B 视觉编码器组成原生多模态模型，256k 上下文窗口、Apache 2.0 协议，这是 Mixtral 之后 Mistral 首个 MoE 旗舰。LMArena 开源非推理赛道直接空降第二，六语言对话能力稳居同类开源第一梯队。\n\n训练侧，Large 3 从零开始用 3000 张 NVIDIA H200 训练；部署侧，Mistral 与 NVIDIA、vLLM、Red Hat 深度协同：FP8 版本可跑在单节点 B200\u002FH200 上，NVFP4 量化版则把门槛压到单节点 8×H100 或 8×A100，Blackwell 专属 attention\u002FMoE kernel、prefill\u002Fdecode 分离推理、投机解码一应俱全。官方明确「推理版本即将到来」，把路线图直接画到了 o3 级选手的对面。\n\n边缘侧同步推出的 Ministral 3 给到 3B \u002F 8B \u002F 14B 三档，每个尺寸都覆盖 base、instruct、reasoning 三个变体，全部内置图像理解，Apache 2.0。Ministral 14B reasoning 在 AIME '25 拿到 85%，而 instruct 版本据说「在多数场景以少一个数量级的 token 数追平或超过同级」，把 token 经济性做成了真正的差异化。\n\nMistral 这次把「前沿 + 开放 + 端侧」三件事压进同一次发布，加上同时登陆 Mistral Studio、AWS Bedrock、Azure Foundry、Hugging Face、Modal、IBM watsonx、OpenRouter、Fireworks、Together、Unsloth AI 等十几家平台，NVIDIA NIM 与 AWS SageMaker 紧随其后。开源协议的边界、推理效率的下沿、硬件覆盖的宽度，这三点 Mistral 一次给齐，留给闭源前沿玩家的护城河又被削掉一层。","https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-3\u002F","2436174c-644b-4a65-9a98-e7a3b705569a",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-06-14T22:02:00Z","2026-06-14T22:10:02.948823Z","2026-06-14T22:10:02.948831Z",true,"agent",5]