[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mistral-large-4-1t-moe":3,"topics-all":38,"news-related-5ddba2a9-781e-4763-b1e0-1e20c6480391":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"5ddba2a9-781e-4763-b1e0-1e20c6480391","Mistral Large 4:1万亿参数MoE,月底开源","Mistral 发布万亿参数 MoE 模型 Large 4,外号 Le Chonk,49B 激活、原生多模态,在欧洲自建的 3800 张 Grace Blackwell 上从零训练。网络安全最亮眼:漏洞复现修补测试拿 82%,Claude Opus 5.5、GPT-6 Astra 因拒绝作答接近零分。权重月底开放。","10 月 6 日,Mistral 上线了公共预览版模型,官方代号很随意——le Chonk,一只胖猫。但参数规模一点不随意:总参数 1 万亿的 MoE,每 token 激活 49B,原生多模态输入。官方承诺权重本月底开放下载([Mistral 官方博客](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-large-4\u002F))。\n\n## 欧洲自建算力,从零训练\n\nML4 没有借别人的云:3,800 张 NVIDIA Grace Blackwell GPU,Mistral 自有欧洲数据中心,从零训练,预览 API 也在同一套设施上。训练数据覆盖 160 余种语言,含欧盟全部官方语言;模型将提供多区域部署,包括一个完全由 Mistral 端到端运营、独立于其他数字服务商、适用欧洲法律的欧洲区域——「AI 主权」是这次发布的核心叙事,也是 €30 亿 D 轮落地后的第一个里程碑。\n\n## 网络安全:闭源模型拒绝回答的地方\n\n最大胆的数字在网络安全:Artificial Analysis Cyber Index 独立评测里全球前五、中国以外开放权重第一;「复现开源软件真实漏洞并修补」测试拿 82%,官方称全场最高;Cybench 40 道安全竞赛题解出 93%。对照组很刺眼:Claude Opus 5.5 和 GPT-6 Astra 在同一测试里接近零分——安全策略让它们拒绝作答。Mistral 的逻辑:防御始于证明漏洞真实存在,而闭源模型的过滤器恰好卡住这一步。红队阶段,他们向安全厂商、审核过的伙伴和政府部门开放了降低审核、扩展网络能力的同款模型。安全侧:Lakera B3 抗提示注入 93.3%,恶意网络请求拒绝率高于所有开源模型。\n\n## 编码与 Agent:对标中国旗舰\n\n编码面:DeepSWE v1.1 得 61.7%、Terminal-Bench 4 28.3%,综合 Coding Agent Index 49.8%,官方称超过 DeepSeek V4 Pro 0813 与 Qwen3.8 Max。Surge AI 盲测里,ML4 以 3.74 排五款模型第二,落后 Claude Opus 5(4.22),压过 Kimi K3(3.59)和 GLM-5.3(3.60)。Agent 面:AutomationBench 657 个业务工作流得 59.9%,官方称超过 Kimi K3、MiMo-V2.6-Pro 和 DeepSeek V4 Pro;视觉定位 Dense 200 上 42% 对 GPT-6-Astra 的 41%。API 定价每百万 token 输入 1.36 美元、输出 4.18 美元。\n\n## 后训练与下一步\n\n后训练同样有信息量:3k GPU 规模下,单次 RL 训练每天产出约 330 亿 token 的 rollout,过滤后约 160 亿可训练;官方称奖励仍在上升、没有饱和迹象。第三方报道称权重预计 10 月 27 日前后公开([TNW](https:\u002F\u002Fthenextweb.com\u002Fnews\u002Fmistral-releases-large-4-a-1-trillion-parameter-open-weight-ai-model))。届时的现实是:1T 总参的 MoE 即便只激活 49B,显存门槛也决定了它主要是企业自部署的货。开放权重的叙事第一次从「追平闭源」转向「做闭源不敢做的事」——网络安全、主权部署、可控审核。对企业,这是欧洲第一个万亿级开放权重选项;对开发者,先等权重落地,再让你的 GPU 说话。","https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-large-4\u002F","2436174c-644b-4a65-9a98-e7a3b705569a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"17095d8c-4b4a-4c36-8bda-ee3f22de13b3","en","Mistral Large 4: A 1T Open-Weight MoE from Europe","Mistral's Large 4: a 1T-param MoE with 49B active, trained in Europe. Cyber test hits 82% where closed models refuse; weights due late October.","On October 6, Mistral launched a public preview of its new model with an informal codename — le Chonk, a fat cat. The spec sheet is anything but casual: a 1-trillion-parameter mixture-of-experts (MoE) model with 49B active parameters per token and natively multimodal input. The company has committed to releasing the open weights by the end of the month ([Mistral's official blog](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-large-4\u002F)).\n\n## Trained from scratch on European soil\n\nML4 does not run on someone else's cloud: it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview API is served on that same infrastructure. Its training data covers more than 160 languages, including every official language of the European Union. The model will be available across multiple regions, including a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law — AI sovereignty is the core narrative of this release, and the first milestone after its €3 billion Series D.\n\n## Cybersecurity: where closed models refuse to answer\n\nThe boldest numbers are in cybersecurity. On the Artificial Analysis Cyber Index, an independent evaluation, ML4 ranks among the top five models globally and leads open-weight models developed outside China. On a test that asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scores 82% — per the official blog, the highest of any model. It also solves 93% of the 40 challenges in Cybench. The most striking contrast: Claude Opus 5.5 and GPT-6 Astra score near zero on the same test because their safety policies make them refuse the task. Mistral's argument is blunt — defending software often starts with proving a flaw is real, exactly the step where closed-model safety filters get in the way. During red-teaming, the company gave cybersecurity leaders, vetted partners, and state authorities access to the same model with reduced moderation and expanded cyber capabilities. On the safety side: 93.3% resistance to indirect prompt injections on Lakera's B3 benchmark, and a refusal rate on malicious cyber prompts higher than all open-source models.\n\n## Coding and agents: benchmarking against Chinese flagships\n\nOn coding: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, and a combined Coding Agent Index of 49.8%, which Mistral says places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind human evaluation by Surge AI, ML4 ranked second of five models at 3.74, behind Claude Opus 5 (4.22) and ahead of Kimi K3 (3.59) and GLM-5.3 (3.60). On the agentic side, it scores 59.9% on AutomationBench's 657 real business workflows, which the company says beats Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro; on visual grounding (Dense 200), it posts 42% against GPT-6-Astra's 41%. API pricing is $1.36 per million input tokens and $4.18 per million output tokens.\n\n## Post-training and what comes next\n\nThe post-training details are informative too. At a scale of 3k GPUs, a single RL training run produces roughly 33 billion tokens of rollouts per day, of which around 16 billion are trainable after filtering. Mistral says training rewards are still climbing with no signs of saturation. Third-party reports suggest the weights will go public around October 27 ([TNW](https:\u002F\u002Fthenextweb.com\u002Fnews\u002Fmistral-releases-large-4-a-1-trillion-parameter-open-weight-ai-model)). The reality then: a 1T-total-parameter MoE, even with only 49B active, has a VRAM footprint that makes it an enterprise self-hosting proposition. For the first time, the open-weight narrative is shifting from \"matching closed models\" to \"doing what closed models won't\" — cybersecurity, sovereign deployment, controllable moderation. For enterprises, this is Europe's first trillion-scale open-weight option; for developers, wait for the weights to land, then let your GPU do the talking.","mistral-large-4-1t-moe","2026-10-06T23:10:00Z","2026-10-06T23:10:16.205522Z","2026-10-06T23:10:16.205530Z",true,"agent",449,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"f3c43720-650b-4054-9349-a1386e06c8fc","Le Chonk 把法国拉回非美\u002F美头部:38 分的 Mistral Large 4","mistral-large-4-le-chonk-intelligence-index-38","2026-10-08T03:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"9fa3427c-cf69-46c7-9720-cd3b646a155b","B站开源35B翻译模型:3B激活,150种语言","bilibili-index-translate-35b-moe","2026-10-04T13:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"0372db20-feaf-42b8-98bd-e42d9c550306","德国Kolibri开源:78B参数只激活3.46B","aleph-alpha-kolibri-1-open-moe","2026-10-03T19:14:02+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"0bc17892-3c47-4081-86a4-3d90afa0c54b","小米 MiMo-V2.6 开源:万亿 MoE 追平 Grok 4.7,Flash 三分之一价格保九成战力","xiaomi-mimo-v2-6-open-weights","2026-09-22T13:02:37+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"21fe3c11-4ba4-4801-b6fc-60c4ae559dc1","Yandex 逆流开源:35B 参数的 T5 MoE,每个 token 只激活 0.6B","yandex-aliceai-t5-sparse-moe","2026-09-16T19:11:43+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"8e730a3d-439b-45cf-961d-f77cf01469fd","Cohere 开源 218B 翻译专用 MoE:25B 激活,自测评分超 DeepL,2×H100 可部署","cohere-north-small-translate","2026-09-11T19:07:20+00:00"]