[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kimi-k3-1-pro-open-weights-huggingface":3,"topics-all":38,"news-related-32560a11-3d59-4397-8f01-e75f680a59db":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},"32560a11-3d59-4397-8f01-e75f680a59db","月之暗面兑现承诺:K3.1 Pro 2.8T 权重开源上架","月之暗面把 Kimi K3.1 Pro 权重上架 Hugging Face,兑现 11 月 4 日前开源的承诺。模型为 2.8T 参数 MoE,每 token 激活 104B,100 万 token 上下文,权重以 MXFP4 发布;许可证在 MIT 骨架上对年营收超 2000 万美元的 MaaS 业务另设条款。","月之暗面 10 月 23 日发布 Kimi K3.1 Pro 时,官方博客留了一句硬承诺:「完整模型权重将在 11 月 4 日前放出」。现在,这批权重已经挂上 Hugging Face 仓库 [moonshotai\u002FKimi-K3.1-Pro](https:\u002F\u002Fhuggingface.co\u002Fmoonshotai\u002FKimi-K3.1-Pro),官方称其为「迄今最强的 3T 级开放权重模型」;第三方追踪站 AI\u002FTLDR 在 11 月 2 日记录到这次上架,距期限还有两天。发布同日官方还上线了 2-3 倍速的 K3.1 Pro Highspeed(API 侧),这次开放的则是标准版 Pro 的完整权重。\n\n## 落地的是一台什么规格的机器\n\n官方模型卡的数字:2.8T 总参数、每 token 激活 104B 的 MoE;93 层中有 69 层 Kimi Delta Attention 和 24 层 Gated DSA,官方说明这一代用 Gated DSA 替换了 K3 的全注意力 Gated MLA 层,换取更高效的长上下文建模。专家侧是每 token 从 896 个专家里选 16 个,外加 2 个共享专家;注意力部分共 96 个头,词表 160K;上下文窗口 1,048,576 token;视觉由 401M 参数的 MoonViT-V2 编码器负责,支持文本加图像输入。对自托管方最实际的一条:权重直接以 MXFP4 格式落盘、激活值 MXFP8,且出自量化感知训练,不是训练后压缩——拿到手的即是可直接部署的低比特成品。模型卡给出 Transformers、vLLM、SGLang、Docker Model Runner 四条部署路径;API 侧在 platform.kimi.ai 选 kimi-k3.1-pro,提供 OpenAI\u002FAnthropic 兼容接口,reasoning_effort 分 low\u002Fhigh\u002Fmax 三档,默认 max。\n\n## 许可证:MIT 骨架,两条商业条款\n\nKimi K3.1 Pro License 的授权段就是标准 MIT 句式:使用、复制、修改、合并、发布、分发、再许可、出售、微调、做衍生品,全部放行。真正的分水岭在后面:其一,若你的 MaaS 业务(让第三方对输入、参数或训练数据拥有实质控制权的推理或微调服务)连同关联方在任意连续 12 个月合计营收超过 2000 万美元,商用前必须与月之暗面另行签约;其二,月活超过 1 亿或月营收超过 2000 万美元的产品,必须在界面显著位置标注「Kimi K3.1-Pro」。内部使用,以及经由官方产品或认证推理伙伴的使用,两条都不适用。\n\n## 官方跑分:赢面和短板都写在表上\n\n以下均为官方模型卡口径。Terminal-Bench 4.0 得 65.7,高于 GPT-6 Astra 的 58.2,略低于 Claude Opus 5.5 的 66.4;HLE-Full(带工具)70.0、BrowseComp 92.1,都是这张表里的最高分;内部基准 Kimi Hallu Bench Hard 上,声称级幻觉率从 K3 的 23.4% 降到 8.3%(官方注明该基准是刻意施压的事实性压力测试,比率不代表日常使用错误率)。短板同样没藏:HLE-Diamond 39.4,远低于 Astra 的 65.8;Program Bench 31.5,对上 Opus 5.5 的 65.0 差了近一半。一张自己标出输赢分布的官方表格,比单边碾压的通稿可信得多。\n\n## 安全自限兑现,独立实测跟上\n\n发布时预告的安全策略在开放权重版里坐实:[官方博客](https:\u002F\u002Fwww.kimi.ai\u002Fblog\u002Fkimi-k3-1-pro)写明,他们在 mid-training 和 RL 两个阶段都识别并移除了网络安全相关数据——理由是漏洞研究能力会随 post-training 规模稳定上升,选择在数据层面设闸,同时保留通用编码任务。独立侧也有旁证:Simon Willison 11 月 3 日用 OpenRouter 上的 kimi-k3.1-pro 跑了「读原始权重猜游戏」实验,9 分 41 秒、19,493 个推理 token 后正确认出 Breakout([实测记录](https:\u002F\u002Fsimonwillison.net\u002F2026\u002FNov\u002F3\u002Fqwen-can-interpret-model-weights\u002F))。\n\n对自托管团队,这道题从此只剩显存预算;对开源圈,「给出日期、按期交付」正在变成旗舰开放权重的信用指标,而 2000 万美元那条线,是月之暗面给「开源」与「商业化」分开标的价——默认自由,规模变现另算。","https:\u002F\u002Fhuggingface.co\u002Fmoonshotai\u002FKimi-K3.1-Pro","0ec8f614-42c7-4256-8591-209e1e39eb6b",[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},"d83b8681-231d-403d-bce2-1596085b2d97","en","Kimi K3.1 Pro open weights land on Hugging Face","Moonshot shipped Kimi K3.1 Pro weights to HF inside its Nov 4 deadline: 2.8T MoE, 104B active, MXFP4, MIT-style license, extra terms for $20M+ MaaS.","When Moonshot AI launched Kimi K3.1 Pro on October 23, the official blog left a hard promise: \"The full model weights will be released by Nov 4.\" Those weights are now live on the Hugging Face repo [moonshotai\u002FKimi-K3.1-Pro](https:\u002F\u002Fhuggingface.co\u002Fmoonshotai\u002FKimi-K3.1-Pro) — a model the company calls its most capable 3T-class open-weight release to date. The tracker AI\u002FTLDR logged the upload on November 2, two days inside the deadline. A 2-3x faster Kimi K3.1 Pro Highspeed shipped on the API side at launch; what opened up now is the full standard-Pro weights.\n\n## The specs that landed\n\nFrom the official model card: a 2.8T-parameter MoE with 104B active per token; 93 layers comprising 69 Kimi Delta Attention layers and 24 Gated DSA layers — Moonshot says Gated DSA replaces the full-attention Gated MLA layers of Kimi K3 for more efficient long-context modeling. Each token routes to 16 of 896 experts plus 2 shared experts; attention runs 96 heads over a 160K vocabulary; the context window is 1,048,576 tokens; vision goes through a 401M-parameter MoonViT-V2 encoder, with text-plus-image input. The most practical detail for self-hosters: weights ship directly in MXFP4 with MXFP8 activations, the product of quantization-aware training rather than post-training compression — what you download is a low-bit artifact ready to deploy. The card lists four deployment paths — Transformers, vLLM, SGLang and Docker Model Runner; on the API side, platform.kimi.ai exposes kimi-k3.1-pro with OpenAI\u002FAnthropic-compatible endpoints, and reasoning_effort supports low\u002Fhigh\u002Fmax, defaulting to max.\n\n## The license: an MIT skeleton with two commercial clauses\n\nThe grant section of the Kimi K3.1 Pro License reads like standard MIT: use, copy, modify, merge, publish, distribute, sublicense, sell, fine-tune and build derivatives — all permitted. The watershed comes later. First, if your Model-as-a-Service business — one that gives third parties meaningful control over inputs, parameters or training data — plus affiliates exceeds 20 million US dollars in aggregate revenue over any consecutive 12 months, you must sign a separate agreement with Moonshot before commercial use. Second, products with more than 100 million monthly active users or more than 20 million dollars in monthly revenue must prominently display \"Kimi K3.1-Pro\" in the user interface. Internal use, and use through Moonshot's official products or certified inference partners, is exempt from both.\n\n## The official benchmark table shows wins and losses\n\nAll numbers below are from the official model card. Terminal-Bench 4.0: 65.7, above GPT-6 Astra's 58.2 and a hair under Claude Opus 5.5's 66.4. HLE-Full with tools: 70.0, and BrowseComp: 92.1 — both the highest scores in the table. On the internal Kimi Hallu Bench Hard, claim-level hallucination fell from 23.4% on K3 to 8.3% (the vendor notes this benchmark deliberately stresses factual accuracy and the rates do not represent everyday error rates). The weak spots are printed too: HLE-Diamond 39.4, far below Astra's 65.8; Program Bench 31.5 against Opus 5.5's 65.0. A vendor table that marks its own losses is worth more than a one-sided press release.\n\n## Safety limits delivered, independent test arrives\n\nThe safety posture previewed at launch is baked into the open weights: the [official blog](https:\u002F\u002Fwww.kimi.ai\u002Fblog\u002Fkimi-k3-1-pro) states Moonshot identified and removed cyber-related data from both the mid-training and RL stages, after observing that vulnerability-research capability rises steadily with post-training scale — the lever is data, while general-purpose coding tasks are retained. There is independent evidence too: on November 3, Simon Willison ran the \"read raw weights, name the game\" experiment through kimi-k3.1-pro on OpenRouter, and it correctly identified Breakout after 9m 41s and 19,493 reasoning tokens ([write-up](https:\u002F\u002Fsimonwillison.net\u002F2026\u002FNov\u002F3\u002Fqwen-can-interpret-model-weights\u002F)).\n\nFor self-hosting teams the question now reduces to a VRAM budget; for the open-source field, \"give a date, ship on it\" is becoming the credit metric of flagship open weights — and the 20-million-dollar line is Moonshot pricing \"open\" and \"commercial\" separately: free by default, pay at scale.","kimi-k3-1-pro-open-weights-huggingface","2026-11-04T23:07:35Z","2026-11-04T23:07:40.198436Z","2026-11-04T23:07:40.198447Z",true,"agent",98,[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},"4d0566f5-ef08-4c5f-befc-09dde8010620","Le Chonk 开源兑现:三周RL让TB冲上41%","mistral-large-4-open-weights-terminal-bench","2026-10-29T17:11:35+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"]