[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mistral-large-4-le-chonk-intelligence-index-38":3,"topics-all":38,"news-related-f3c43720-650b-4054-9349-a1386e06c8fc":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},"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 预览版在 Artificial Analysis 智能指数拿 38 分,与 GPT-6 Luna、DeepSeek V4.1 Flash 同档,法国再成美中之外最强模型产地。1T 总参\u002F49B 激活 MoE,网络防御指数 50,首两周半价单任务 0.57 美元,完整权重月底开源。","Artificial Analysis 给 Mistral Large 4 的智能指数打出 38 分,与 GPT-6 Luna(max)、DeepSeek V4.1 Flash(max)同档,法国再次坐到「美中之外最强模型」的位置。38 分把韩国、阿联酋系模型压在身后;Le Chonk 这只「胖猫」是 Mistral 用 3800 张 Grace Blackwell 在欧洲自建数据中心从零训出的。\n\n## 智能指数 38 分代表什么\n\n按 Artificial Analysis 口径,Intelligence Index 把 reasoning、agentic、coding、知识任务折算成单一可比分数。Le Chonk 的 38 落在开源权重模型最拥挤的区间:DeepSeek V4.1 Flash(max) 39、GLM-5.3-Flash 36 附近、Qwen3.8 Max 与 Kimi K3 在 35-37 之间。Mistral 把多模态输入门槛放宽到单请求 100 张图(老版 8 张)是这条曲线的关键抬升器:文档图表、卫星图、工程图纸这类「图多于文」的输入,Le Chonk 不再需要切片拼装。\n\n智能指数之外,Artificial Analysis 网络防御指数(Cyber Index)给出 50,和 GLM-5.3-Flash 平手,排在 MiMo-V2.6-Pro(56)之后。子项 CyberGym-E2E-AA 上 Le Chonk 拿到 82%,比 MiMo-V2.6-Pro(79%)、GPT-6 Luna(max, 78%)都高——这是开源权重模型不再被「拒答」卡住的特殊战场。\n\n## 成本曲线在「头两周半价」之外仍然硬\n\nMistral 把单任务成本摆到对比图横轴,标价 $1.13 \u002F 任务,首两周半价期间降到 $0.56 \u002F 任务。对比 GLM-5.3-Flash($0.25)、DeepSeek V4.1 Flash(max, $0.27)仍是 4 倍级别。Mistral 的解释是首两周折扣「让体验铺出去」,权重开放后自有部署成本回到硬件层而非 API 标价层。对欧洲客户来说这条数字才合规:跑在自己机房里,token 不出数据中心,账单按算力走。\n\n标准价 $1.36 \u002F $4.18 每百万 token($0.14 \u002F 缓存输入),512k 上下文,文本与图像进、文本出。Mistral 在欧洲运营独立,符合「主权 AI」叙事,但代价是同样智能指数下,token 单价比中国开源权重模型贵一截。\n\n## 「不是中国也不是美国」本身就是一道保险\n\nLe Chonk 不是技术突破——38 分在 2026 年 10 月不稀奇——它真正的新闻点在于供应源分散。法国 + 欧洲自建数据中心 + 完整权重月底落地 + 主权叙事,是一组绑定方案,直接回应企业「不敢用中国模型、也不敢只用美国模型」的两难。Microsoft-Mistral 几周前签下的 Vera Rubin GPU 主权云协议是同条曲线的另一端,Artificial Analysis 这条报道是在给这条曲线贴价格标签。\n\n月底开源后,如果社区在 Mistral 训练配方上跑出微调,把 38 分推到 40-42 区间,Le Chonk 就会成为欧洲开源权重模型的新默认基线。在那之前,38 分这件事本身,和韩国 Solar Mini 4(24 分)、阿联酋系模型同档并列,已经够让「法国的 LLM 又回来了」成为今年的副标题之一。","https:\u002F\u002Fartificialanalysis.ai\u002Farticles\u002Fmistral-large-4-france-ai","3d6c2ae2-a449-467d-91dd-68fbbd04d714",[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},"4345034b-90b8-4768-8341-59f12946c94c","en","Le Chonk: Mistral Large 4 Hits Intelligence Index 38","Mistral Large 4 preview scores 38 on the Artificial Analysis Intelligence Index, level with GPT-6 Luna (max) and DeepSeek V4.1 Flash (max). France reclaims strongest-model-from-outside-US-China status. 1T params \u002F 49B active MoE. Cyber Index 50. Weights end of October.","Artificial Analysis has given Mistral Large 4 an Intelligence Index score of 38, putting it level with GPT-6 Luna (max) and DeepSeek V4.1 Flash (max). France is back in the seat of \"the strongest model from outside the US and China.\" The 38-point mark puts Korean and UAE-developed models in their place. Le Chonk (a deliberately self-deprecating nickname, French for \"the chonk\") was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers.\n\n## What the Intelligence Index 38 actually means\n\nUnder Artificial Analysis's methodology, the Intelligence Index folds reasoning, agentic, coding, and knowledge tasks into a single comparable score. Le Chonk's 38 sits in the most crowded band of open-weight models: DeepSeek V4.1 Flash (max) is 39, GLM-5.3-Flash sits around 36, and Qwen3.8 Max and Kimi K3 land in the 35-37 range. Compared with Mistral Large 3 on the same methodology, the new model lifts the score by 12 points. Mistral raised the multimodal input ceiling to 100 images per request (up from 8 on older Mistral models) — that change is the key driver on this curve. Document charts, satellite imagery, and engineering drawings — inputs where images outnumber words — no longer need to be sliced and reassembled.\n\nBeyond the Intelligence Index, Artificial Analysis's Cyber Index gave 50, tied with GLM-5.3-Flash and behind MiMo-V2.6-Pro (56). On the CyberGym-E2E-AA sub-test, Le Chonk scored 82%, ahead of MiMo-V2.6-Pro (79%) and GPT-6 Luna (max, 78%). This is the specific field where open-weight models stop being blocked by refusals.\n\n## The cost curve is still hard beyond the \"two-week half-price\" window\n\nMistral puts single-task cost on the comparison chart's x-axis. List price is $1.13 per task; during the first two weeks, the launch discount brings it down to $0.56 per task. Against GLM-5.3-Flash ($0.25) and DeepSeek V4.1 Flash (max, $0.27), that's still roughly 4x more costly. Their explanation: the first two weeks are for getting hands-on time, and after the weights open, self-hosted deployment costs drop back to the hardware layer rather than the API sticker layer. For European customers, this is the number that actually matters for compliance: running on their own infrastructure, tokens never leave the data center, and the bill is on compute rather than per token.\n\nStandard pricing is $1.36 \u002F $4.18 per million input\u002Foutput tokens ($0.14 per million cached input), 512k context, text and image in, text out. Mistral operates independently in Europe, which fits the \"sovereign AI\" narrative. The cost is that, at the same Intelligence Index score, the per-token price is still a step above Chinese open-weight models.\n\n## \"Not China and not the US\" is itself a kind of insurance\n\nLe Chonk is not a technical breakthrough — 38 on the Intelligence Index is not rare in October 2026 — the actual news angle is supply-source diversification. France plus European self-built data centers plus full open weights landing at month-end plus a sovereignty narrative is a packaged answer to enterprises caught between \"can't use Chinese models\" and \"can't use only American models.\" Microsoft-Mistral's Vera Rubin GPU sovereign cloud agreement from a few weeks back is the other end of the same curve. The Artificial Analysis piece is putting a price tag on that curve.\n\nAfter end-of-month open-weights, if the community runs further fine-tuning on Mistral's training recipe and pushes 38 into the 40-42 range, Le Chonk becomes the new default baseline for European open-weight models. Until then, the 38 itself, alongside Korean Solar Mini 4 (24) and UAE-developed models in the same band, is enough to put \"France's LLM is back\" on this year's list of side headlines.","mistral-large-4-le-chonk-intelligence-index-38","2026-10-08T03:30:00Z","2026-10-08T03:17:25.405304Z","2026-10-08T03:17:25.405314Z",true,"agent",303,[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},"5ddba2a9-781e-4763-b1e0-1e20c6480391","Mistral Large 4:1万亿参数MoE,月底开源","mistral-large-4-1t-moe","2026-10-06T23:10:00+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"]