[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mistral-large-4-open-weights-terminal-bench":3,"topics-all":38,"news-related-4d0566f5-ef08-4c5f-befc-09dde8010620":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},"4d0566f5-ef08-4c5f-befc-09dde8010620","Le Chonk 开源兑现:三周RL让TB冲上41%","Mistral 如约放出 Le Chonk 开源权重:1T 参数多模态 MoE 登陆 Hugging Face。三周强化学习把 Terminal-Bench 4.0 从 28.3% 抬到 41%,输出 token 省三成;编码智能体指数 48% 追平 GLM-5.3,每任务少用 47% token。","10 月 6 日,Mistral 把旗舰 Mistral Large 4(Le Chonk)放进公开预览,留下一句「权重月底开放」。10 月 29 日,权重如约登上 Hugging Face——且已不是预览那个版本:官方称这三周强化学习一直在跑,放出来的 checkpoint 比预览版强出一截([官方公告](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-large-4-oss\u002F))。\n\n## 三周 RL,数字全线上抬\n\n最直观的是 Terminal-Bench 4.0:预览期 28.3%,开源版 41%,官方口径「提升 13 个百分点」,输出 token 还少了 30%。官方对比图里,最接近的 GLM-5.3 为 40%、DeepSeek V4.1 Flash 1017 为 31%。Cognition 私有基准 FrontierCode 2.0 上,它以 19.7% 领先 GLM-5.3(17.8%)、Kimi K3(17.3%)与 DeepSeek V4.1 Flash(19.3%)。\n\nArtificial Analysis 编码智能体指数拿到 48%,与 GLM-5.3 打平,平均每任务却少用 47% 的 token——Mistral 的卖点正在这:分数追平,成本打下来。其他维度:BrowseComp 84.5%;金融基准 Finch 73.3%,开源仅次于 Kimi K3;B3 抗攻击率 97.1%。推理档位改为 low\u002Fmedium\u002Fhigh 三档,默认 high。\n\n## 1T 参数的部署门槛\n\n架构与预览期一致:总参约 1T,每 token 激活 49B(算上嵌入与输出层 52B),128 个路由专家选 4 个、外加 1 个共享专家;注意力用 MLA 加 DeepSeek Sparse Attention 索引器,上下文 1M,视觉走 1.6B 的 Pixtral 编码器。官方放出 FP8\u002FNVFP4 两版,llama.cpp 与 Unsloth 同日跟上 GGUF。\n\nHugging Face 官方博客实测:2-bit 量化版在单台 512GB 统一内存的 Apple Silicon Ultra 上即可跑,M3 Ultra 上 Q2_K 预填约 90 token\u002Fs、生成约 15 token\u002Fs;vLLM 跑 NVFP4 需要 Blackwell 显卡。许可证是修改版 MIT「Mistral Ecosystem Open-Weight License」:使用、微调、商用免费,唯一条件是做模型即服务(MaaS)需另签商业授权。API 每百万 token 输入 0.70 美元、缓存命中 0.07 美元、输出 2.10 美元。\n\n## 两点冷思考\n\n其一,「编码最强开源」是 Mistral 自己的口径:TB 与 FrontierCode 的领先只在官方图表里成立,AA Intelligence Index v5 上它仅 26 分——西方开放权重最高,开源阵营里只排第五,综合榜前排仍是中国模型。其二,网安叙事要拆开看:整体 Cyber Index 51%,低于 DeepSeek V4.1 Flash 的 57% 与 MiMo-V2.6-Pro 的 56%;真正差异点是「防守不拒绝」:漏洞复现这类正当防御不被安全滤镜拦下,恶意网安请求平均拒绝率却有 96.6%,配上 Booz Allen、Cloudflare 等厂商背书与欧洲主权叙事。\n\n这次发布真正的信号,不是又一个万亿参数开源,而是开源旗舰从「发布即巅峰」变成「发布后还在长」,增量恰好长在 token 效率与部署门槛这两个直接决定成本的地方。盯成本的团队不妨拿自己的任务集跑一轮——追平 GLM-5.3 还省 47% token,值得进候选池。","https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-large-4-oss\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},"3b1067f9-02d3-4ce5-8ddd-e79f1217ee1f","en","Mistral Large 4 weights land: Terminal-Bench 4.0 hits 41%","Mistral Large 4 open weights: TB 4.0 hits 41% after three RL weeks; ties GLM-5.3 at 48% on the Coding Agent Index with 47% fewer tokens.","When Mistral put its flagship Mistral Large 4 (Le Chonk) into public preview on October 6, it promised open weights by the end of the month. On October 29 those weights landed on Hugging Face — and they are not the preview checkpoint: Mistral says the reinforcement-learning run kept going for three more weeks, and the released model is meaningfully stronger ([official post](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-large-4-oss\u002F)).\n\n## Three more RL weeks, numbers up across the board\n\nTerminal-Bench 4.0 is the headline: 28.3% at preview, 41% at release — \"a 13-percentage-point improvement\" in Mistral's words — while output tokens dropped 30%. In Mistral's comparison chart, GLM-5.3 sits at 40%, DeepSeek V4.1 Flash 1017 at 31%, Kimi K3 at 21%. On Cognition's private FrontierCode 2.0, Le Chonk leads with 19.7%, ahead of Kimi K3 (17.3%), GLM-5.3 (17.8%) and DeepSeek V4.1 Flash (19.3%).\n\nOn the Artificial Analysis Coding Agent Index the new model scores 48%, tied with GLM-5.3, while using 47% fewer tokens per task — the efficiency angle Mistral keeps pushing: match the score, cut the inference bill. Elsewhere: BrowseComp 84.5%; 73.3% on the finance benchmark Finch (second among open models, behind Kimi K3); and 97.1% attack resistance on the B3 agent-security benchmark. Reasoning effort now ships as low\u002Fmedium\u002Fhigh, with high as default.\n\n## What a 1T model costs to run\n\nThe architecture matches the preview: about 1T total parameters, 49B active per token (52B including embeddings and output layers), 128 routed experts with 4 active per token plus one shared expert, MLA attention with a DeepSeek Sparse Attention indexer, a 1M context window, and a 1.6B Pixtral vision encoder. Official checkpoints ship in FP8 and NVFP4, with day-one GGUF quants from llama.cpp and Unsloth.\n\nHugging Face ran its own hardware tests: the 2-bit quants fit on a single Apple Silicon Ultra machine with 512GB of unified RAM — on an M3 Ultra, Q2_K prefills at about 90 tok\u002Fs and generates around 15 tok\u002Fs; vLLM with the NVFP4 checkpoint requires Blackwell GPUs. The license is a modified MIT (\"Mistral Ecosystem Open-Weight License\"): free to use, modify, fine-tune and commercialize, except that Model-as-a-Service businesses need a separate commercial license. API pricing is $0.70\u002FM input, $0.07\u002FM cached input and $2.10\u002FM output.\n\n## Two caveats\n\nFirst, \"best open-weight model for coding\" is Mistral's own framing: the Terminal-Bench and FrontierCode leads hold in its charts, but on the AA Intelligence Index v5 it scores 26 — the highest among Western open-weight models, yet only fifth among open weights overall, with Chinese models ahead on the aggregate. Second, the cybersecurity story needs unpacking: its 51% on the overall Cyber Index trails DeepSeek V4.1 Flash (57%) and MiMo-V2.6-Pro (56%). The real differentiator is \"defense without refusals\": legitimate reproduce-and-patch work does not trip safety filters, malicious cyber requests still get refused 96.6% of the time, and the release comes with endorsements from Booz Allen, Cloudflare and Orange Cyberdefense plus a European-sovereignty pitch.\n\nThe signal worth keeping: an open-weight flagship that keeps improving after its announcement, with the gains landing exactly where they hit your bill — token efficiency and deployment threshold. If you track inference costs, a model that ties GLM-5.3 at 48% while spending 47% fewer tokens deserves a run through your own evals.","mistral-large-4-open-weights-terminal-bench","2026-10-29T17:11:35Z","2026-10-29T17:11:47.825591Z","2026-10-29T17:11:47.825600Z",true,"agent",395,[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},"32560a11-3d59-4397-8f01-e75f680a59db","月之暗面兑现承诺:K3.1 Pro 2.8T 权重开源上架","kimi-k3-1-pro-open-weights-huggingface","2026-11-04T23:07:35+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"03d33c23-e56a-41eb-8732-b0712059662f","PUBG 母公司开源 MoE:激活仅 4.3B 单卡可跑","krafton-raon-1-mini-preview-open-weights","2026-10-23T13:08:33+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"cee60a45-f212-4a78-83dd-0226c07b3460","腾讯开源记忆写入模型 Hy-Memory:30B MoE 只激活 3B,四框架平均分超 Opus 5","tencent-hy-memory-30b-memory-writer","2026-10-17T15:14:21+00:00"]