[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mozilla-open-weights-4-month-gap":3,"topics-all":38,"news-related-b7668f43-05a7-46d7-845f-27e70fcaceec":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},"b7668f43-05a7-46d7-845f-27e70fcaceec","Mozilla 报告:中国开放权重距美国前沿模型只差 4.4 个月","Mozilla《State of Open Source AI》v1.1 用 METR 数据拟合:中国最好的开放权重模型距美国闭源前沿只剩约 4.4 个月,Epoch AI 独立估计 4 个月;Kimi K3 距 Claude Fable 5 仅 2.1 分、价格约 30%。报告建议多数组织默认用开放模型。","9 月 15 日,Mozilla 发布了《State of Open Source AI》报告 v1.1(首版 7 月 14 日上线,本版数据截至 9 月 1 日)。核心数字:美国闭源前沿模型与中国最好的开放权重模型之间,能力差距只剩约 4.4 个月——这是 Mozilla 用 METR Time Horizon 数据拟合的结果,独立机构 Epoch AI 的估计是 4 个月,两个数字互相咬合。\n\n## 「代差」已经变成「季度差」\n\n报告引用 Artificial Analysis 智能指数 v4.1.1(9 月 1 日数据):前十名里,前四席是闭源(Claude Opus 5、Claude Fable 5、GPT-5.6 Sol、Grok 4.6),紧接着的四个席位全部是开放权重——Kimi K3 与 GLM-5.3 并列 60.0 分,距第一名 Claude Opus 5 的 63.0 分只差 3 分,距 Claude Fable 5 的 62.1 分只差 2.1 分。\n\n差 2.1 分意味着什么?报告给出价格对照:Kimi K3 API 定价 3\u002F15 美元(输入\u002F输出每百万 token),Claude Fable 5 是 10\u002F50 美元——约 30% 的价格。用三成价钱买 2 分的差距,这笔账任何人都会算。\n\n## 闭源前沿真正守住的三个阵地\n\n报告没有回避闭源的剩余优势。CTO Raffi Krikorian 点名三类闭源仍值得溢价的任务:专家级专业工作、高强度检索、长上下文。硬数字也支撑这个判断:GDPval-AA v2 专家知识工作基准上,Fable 5 领先 K3 达 92 Elo,是两者共享基准里最大的分差;100 万 token 多针检索,Gemini 3.1 Pro 拿到 89%,DeepSeek V4-Pro 只有 41%;METR 时间线数据下,8 小时以内的任务开源闭源都能胜任,超过 12 小时的谁都不行——分界带就在中间这段。\n\n「开放默认、闭源按需」已经有现实模板:报告引用 DoorDash 的做法——日常工作量跑 Kimi,复杂任务留给 Fable。\n\n## 冷水:能力追平了,钱包和部署没有\n\n报告里最扎心的对比是收入:2025 年模型层收入 96% 归闭源厂商,开放模型只拿 4%;而同期闭源模型在约 90% 能力对齐下,每次调用贵约 6 倍。投产率同样有差距:调查显示运行开放模型的团队只有 53% 真正投产,闭源团队是 63%。\n\n但天平在动:2026 年 8 月是开放模型第一次在 OpenRouter 按请求数登顶的月份,按 token 量计的前十模型里 8 个是开放权重、其中 7 个来自中国厂商;美国闭源厂商的请求份额从 5 月的 54% 掉到 8 月底的 44%。还有一个值得玩味的信号:8 月 16 日 DeepSeek 完成开放模型第一次官方涨价,输出价格上调 2.3–4.6 倍——当开放模型也握有定价权,「便宜」这个标签还能贴多久,是个真问题。\n\n## 所以呢\n\n这份报告真正的读者不是模型厂商,而是每季度都要写 AI 预算的技术负责人:你在为「4 个月的领先」支付几倍的溢价,而你的日常工作负载里,有多少真的落在这 4 个月里?报告给多数组织的答案很直接——默认选项换成开放权重,把闭源前沿留给那一小段真正需要它的任务。\n\n参考:[Mozilla《State of Open Source AI》v1.1](https:\u002F\u002Fstateofopensource.ai\u002F)、Tom's Hardware(tomshardware.com)与 Ars Technica 9 月 15 日报道","https:\u002F\u002Fstateofopensource.ai\u002F","8a4f9b4d-8163-4866-93d0-3611ef77a0bb",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"8ddf2b28-0234-41a4-9862-3f0faef96472","market-analysis",{"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},"743a6637-842d-417d-9009-7a744748b3a6","en","Mozilla: Chinese Open Weights 4.4 Months Behind US Frontier","Mozilla's report puts the best Chinese open-weight models 4.4 months behind US closed frontier; Kimi K3 trails Fable 5 by 2.1 points at 30% of the price.","On September 15 Mozilla published v1.1 of its State of Open Source AI report (first edition July 14, data current to September 1). The headline number: the capability gap between US closed frontier models and the best Chinese open-weight models has narrowed to roughly 4.4 months — Mozilla's own fit on METR time-horizon data, cross-checked against Epoch AI's independent estimate of four months.\n\n## From a generation gap to a quarterly one\n\nThe report cites the Artificial Analysis Intelligence Index v4.1.1 (September 1 data): in the top ten, the first four seats are closed (Claude Opus 5, Claude Fable 5, GPT-5.6 Sol, Grok 4.6), and the next four are all open weights — Kimi K3 and GLM-5.3 tied at 60.0, three points behind leader Claude Opus 5 at 63.0 and 2.1 points behind Claude Fable 5 at 62.1.\n\nThe price comparison is the punchline: Kimi K3 lists at 3\u002F15 USD per million input\u002Foutput tokens against Fable 5's 10\u002F50 USD — roughly 30% of the cost for a two-point deficit.\n\n## Where closed models still earn their premium\n\nThe report is candid about what closed frontier still buys. CTO Raffi Krikorian names three areas: expert professional work, high-intensity retrieval, and long context. The hard numbers back him up: Fable 5 leads K3 by 92 Elo on GDPval-AA v2, the largest gap among their shared benchmarks; on 1M-token multi-needle retrieval, Gemini 3.1 Pro scores 89% against DeepSeek V4-Pro's 41%; and on METR's time-horizon data, both open and closed models handle tasks under eight hours, while past twelve hours neither does.\n\nThe open-default, closed-on-demand pattern already has a template: DoorDash, cited in the report, runs Kimi for routine workloads and reserves Fable for harder tasks.\n\n## Cold water: capability parity, not wallet parity\n\nThe sharpest contrast in the report is revenue: in 2025, closed providers captured 96% of model-layer revenue while open models took 4% — even as closed models cost roughly 6x more per call at around 90% parity. Production rates lag too: only 53% of teams running open models get them into production, versus 63% for closed-model teams.\n\nBut the balance is moving. August 2026 was the first month an open model led OpenRouter by request count; eight of the top ten models by token volume were open weights, seven of them Chinese-built; US closed providers' request share fell from 54% in May to 44% by late August. One more telling signal: on August 16, DeepSeek delivered the first list-price increase by an open model, raising output prices 2.3-4.6x — when open models gain pricing power, the cheap label deserves a second look.\n\n## So what\n\nThe real audience for this report is not model vendors but every engineering leader writing an AI budget this quarter: how many of your workloads actually fall inside that four-month head start, and what premium are you paying to cover them? For most organizations the answer is blunt — default to open weights, and reserve closed frontier for the narrow band of tasks that genuinely needs it.\n\nReferences: [Mozilla, State of Open Source AI v1.1](https:\u002F\u002Fstateofopensource.ai\u002F); [Tom's Hardware coverage](https:\u002F\u002Fwww.tomshardware.com\u002Ftech-industry\u002Fartificial-intelligence\u002Fchinas-open-weight-ai-models-are-now-just-4-months-behind-frontier-us-offerings-mozilla-report-claims-models-still-lag-in-some-benchmarks-but-are-drastically-cheaper-to-use)","mozilla-open-weights-4-month-gap","2026-09-16T13:08:54Z","2026-09-16T13:09:15.385586Z","2026-09-16T13:09:15.385598Z",true,"agent",95,[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},"8d7b30e0-996f-4141-8501-8f464bda6282","中美开放权重参数上限差距拉到 20 倍:Hugging Face 夏季报告里的三条隐藏数据","hugging-face-summer-2026-frontier-ceiling","2026-08-22T14:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"ad3be632-49a1-44f5-9816-62c168e56467","全球大模型调用量榜前五全是\"中国造\":开源 MoE 正在重写 OpenRouter 的地理坐标","openrouter-top5-china-moe-open-source-2026w31","2026-08-02T03:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"e84fe968-5d86-4247-baad-5da23efef860","UltraData-RL-2609 开源:85,995 条可验证奖励任务,拆解 MiniCPM5-2B 的 RL 燃料","ultradata-rl-2609-verifiable-rl-dataset","2026-09-07T23:07:45+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"2b37a19b-1dde-4238-bef5-39b1d19157f1","OpenBMB 开源 MiniCPM5-2B:2B 端侧模型平均分超对比集 4B 级","openbmb-minicpm5-2b-on-device","2026-09-07T17:02:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"453ce9a1-5d55-4981-b44d-c261b8051724","GLM-5.3 753B 权重上架 HuggingFace,智谱兑现两周开源承诺","glm-5-3-weights-huggingface-release","2026-08-28T15:15:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5a6c3aea-9376-4a3d-a79f-a0d19406e299","月之暗面 Kimi K3 想从微软、AWS、Google 手里分一杯羹：中美大模型的收益分成时代","moonshot-kimi-k3-cloud-revenue-share-talks","2026-08-28T12:30:00+00:00"]