[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hugging-face-summer-2026-frontier-ceiling":3,"news-related-8d7b30e0-996f-4141-8501-8f464bda6282":38},{"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},"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 夏季报告里藏了三条被多数报道忽略的事实:中国 AI 实验室月月发模型,参数上限 754B-2.78T;美国厂商除 NVIDIA\u002FAMD 多在 130B 以下;Inkling 这种美国前沿模型实际基于中国模型构建。","Hugging Face 8 月 14 日发的《State of Open Models: Summer 2026》里讨论最多的是 Qwen 拿下 15 万衍生模型、GGUF 仓库一年涨 464%。但\"The frontier is moving fast\"一节藏着三条被多数转载忽略的结构性数据。\n\n## 中国前沿参数天花板:754B 到 2.78T\n\n2026 年几乎每个月,中国 AI 实验室发的\"最大且最强\"开放权重模型,参数都大于同期美国实验室的任意模型。中国侧月度上限跑在 754B 到 2.78T 之间;美国侧七个月里有五个月停在 130B 以下。少数例外:NVIDIA 5、6 月发了 561B 的 Nemotron 3 Ultra;Thinking Machines 的 Inkling 是另一个被点名的\"美国前沿\"。\n\n报告给了一个刺眼脚注:在 100B 以上尺度,部分美国前沿模型基于中国模型构建或复用中国实验室产物,Inkling(952B)正是被列入此类。\n\n## 美国前沿的中国血统\n\n报告原文:\"At the frontier scale, some U.S. model releases above 100B parameters this year are built on top of Chinese models or leverage artifacts from Chinese labs, such as Thinking Machines' Inkling (952B).\"报告列出的美国原创前沿数字:NVIDIA Nemotron 3 Ultra(561B)、Nemotron 3 Super(124B)、Arcee AI Trinity-Large(399B)。\n\n这条事实解释了为什么美国头部厂商只能在 130B-561B 区间出货,却拿不出对标 Qwen 3.8 Max(2.4T)规模的开源模型。Inkling 能撞到 952B,基座源自中国体系——\"Inkling 是美国模型\"的常识需要重新审视。\n\n## NVIDIA 和 AMD 成开放权重推手\n\n2026 年发新开放权重模型最多的两家既不是 OpenAI 也不是 Anthropic,而是 AMD 和 NVIDIA,各自超过 200 个新模型仓库。报告解释得很功利:开放模型是卖芯片的手段——一个针对你硬件优化、又免费的模型,是硬件能跑的最好证明。AMD 集中在\"转化层\",把万亿参数模型移植到美国硬件栈上跑通;NVIDIA 走 Nemotron 全家桶。这跟中国实验室围绕国产硬件做优化形成镜像。\n\n## 许可证的\"反常识\"分布\n\n第 3 节还有一个对比被多数转载忽略:178 个中国发的 200B+ 参数模型中,59% 用 Apache 2.0,22% 用 MIT,几乎没有任何非商业限制;同尺寸美国侧只有 29% 是 Apache 或 MIT,41% 是定制条款,30% 干脆没声明许可证。Kimi K3 与 Qwen 3.8 2.4T 加入非商业限制和收入分成条款,开源权重不再等于\"放弃商业回报\"——这套叙事还能跑多久,是这份报告留给 2026 下半年的真问题。\n\n数据来源:Hugging Face 报告原文 (https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026),Solidot 8 月 17 日转引(https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85118)。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026#the-frontier-is-moving-fast","24d5c6c5-6573-4180-a1fd-f1459842d1af",[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},"23959fdd-5b63-4c28-a07f-7c0cc736c574","en","Hugging Face Summer 2026 report: three hidden data points behind the US-China open-weight frontier ceiling","Hugging Face Summer 2026 report hides three overlooked structural facts: Chinese labs ship largest open models monthly with a 754B-2.78T parameter ceiling; US labs except NVIDIA\u002FAMD stay below 130B; US frontier Inkling is built on top of Chinese models.","Hugging Face published State of Open Models: Summer 2026 on August 14. Most coverage zoomed in on Qwen grabbing 151,448 derivatives and GGUF repositories rising 464% year-on-year. The frontier ceiling section, however, contains three structural data points that most retellings skipped.\n\n## Chinese frontier ceiling: 754B to 2.78T parameters\n\nIn almost every month of 2026, the largest and most performant open-weight model from a Chinese lab was larger than any model an American lab released. China monthly ceiling: 754B to 2.78T parameters. US: stayed below 130B in five of seven months. Exceptions: NVIDIA Nemotron 3 Ultra (561B) in May and June, and Thinking Machines Inkling as the other named \"US frontier.\"\n\nThe report then drops a footnote that cuts the other way: at the 100B-and-above scale, some US frontier models are built on top of Chinese models or reuse artifacts from Chinese labs. Inkling (952B) lands squarely in that category.\n\n## The Chinese backbone of US frontier models\n\nThe original text: \"At the frontier scale, some U.S. model releases above 100B parameters this year are built on top of Chinese models or leverage artifacts from Chinese labs, such as Thinking Machines' Inkling (952B).\" The original US frontier numbers listed: NVIDIA Nemotron 3 Ultra (561B), Nemotron 3 Super (124B), Arcee AI Trinity-Large (399B).\n\nThat fact explains why US vendors can ship between 130B and 561B but nothing to match Qwen 3.8 Max (2.4T). Inkling reaches 952B because its base traces back to a Chinese stack. The conventional framing of Inkling as a US model needs to be re-examined.\n\n## NVIDIA and AMD as the de facto open-weight publishers\n\nIn 2026 the two organizations publishing the most new open-weight models are neither OpenAI nor Anthropic, but AMD and NVIDIA, each with more than 200 new model repositories. LiquidAI ranks third at around 100.\n\nThe report explains this in unusually frank terms: open models are a way to sell chips. A model optimized for your hardware and freely available is the clearest proof that the hardware works. AMD contributes at the conversion layer, porting trillion-parameter models onto US hardware stacks. NVIDIA runs the Nemotron family. The pattern mirrors what Chinese labs do when they tune models for domestic chips. Both sides are using open weight as ammunition in the chip war.\n\n## The reverse-rational license distribution\n\nSection 3 hides another contrast most retellings ignored: of 178 Chinese releases above 20B parameters, 59% carry Apache 2.0 and 22% carry MIT, almost none with non-commercial restrictions. On the US side of the same size band, only 29% is Apache or MIT, 41% sits under custom terms, 30% declares nothing at all.\n\nCombined with the recent move by Kimi K3 and Qwen 3.8 2.4T to add non-commercial restrictions and revenue-share requirements, the takeaway is this: open weight no longer means giving up commercial returns. API, cloud, and ecosystem position are the real business. But when the flagship open-weight models themselves start adding revenue-share clauses, how long can that narrative hold up? That is the question the report leaves for the second half of 2026.\n\nData sources: Hugging Face report (https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026), Solidot reprint August 17 (https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85118).","hugging-face-summer-2026-frontier-ceiling","2026-08-22T14:00:00Z","2026-08-22T05:10:00.203886Z","2026-08-22T05:10:00.203898Z",true,"agent",68,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"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":46,"title":47,"news_slug":48,"published_at":49},"f6e4aab0-7693-4c2c-bb66-c1641fc2cc3e","Ox Alpha 谜底揭晓:智谱 GLM-5.3-Flash,MIT 开源 320B MoE","ox-alpha-glm-5-3-flash-reveal","2026-08-27T13:30:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"68072ee1-fc37-4064-ab18-09550ae72d1b","GLM-5.3-Flash 把 320B MoE 跑在国产芯片上:Flash 价位和 $0.15 API 的混合注意力栈","glm-5-3-flash-chinese-chips-hybrid-attention","2026-08-27T03:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"0d8fdf45-4585-47c0-9e78-3652e318b156","Apple Intelligence 中国版落地:通义千问接管语言 AI,百度负责视觉搜索","apple-intelligence-china-qwen-baidu-2026","2026-08-25T12:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"b863d01a-dfdf-41c9-8266-e4602e58bde3","Qwen3.8-Max 开源权重落地:砍掉视觉与 1M 上下文,许可证换成收入分成","qwen3-8-max-open-weights-stripped-relicense","2026-08-23T13:30:00+00:00"]