[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hf-summer-2026-china-open-weight-parameter-ceiling":3,"news-related-9389d1ed-dd2d-41cb-bbc5-9a543e2b2f71":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},"9389d1ed-dd2d-41cb-bbc5-9a543e2b2f71","开源报告里的「参数天花板」分水岭:中国实验室把上限拉到2.78T,美国还在130B徘徊","Hugging Face 夏季报告:几乎每个月中国前沿实验室的最大模型参数都在 754B–2.78T 区间,美国 7 个月里有 5 个月没摸到 130B。Nemotron 3 Ultra(561B)与 Inkling(952B)是仅有的例外,后者明确「依托中国模型构建」。","## 一个被「量化层」掩盖的数字\n\n8 月 14 日,Hugging Face 发布《State of Open Models: Summer 2026 Observations》报告,把 2026 年 1–8 月 Hub 上的开源权重生态盘了一遍。比起广为流传的「Qwen 拿下 15 万衍生模型」与「GGUF 仓库涨 464%」,报告第一章节「The frontier is moving fast」里有一张不那么显眼、但杀伤力更大的图:**Largest open-model releases from Chinese and US labs by month in 2026**。这张图把「开源」这个词的地缘重量第一次用纯数字钉在了桌面上。\n\n## 「参数天花板」的真实差距\n\n报告给出了一个极其简洁的事实:**In almost every month of 2026, the largest and most performant open model from a Chinese lab was larger than any model an American lab released.** 具体数字是:中国前沿实验室 2026 年每个月的最大开源模型,参数规模落在 **754B 到 2.78 万亿** 之间;而美国实验室 7 个月里有 **5 个月** 没摸到 130B。\n\n唯一两个打破这一格局的美国开源旗舰是 NVIDIA 的 **Nemotron 3 Ultra**(2026 年 5 月与 6 月发布,561B 总参 \u002F 55B 激活)与 Thinking Machines Lab 的 **Inkling**(952B)。报告原文里,Nemotron 3 Ultra 是和 Inkling 一起被点名的「例外」,但接下来还有一段更具冲击力的补充:**「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)」**。换句话说,美国 100B 以上的开源旗舰里相当一部分「依托中国模型构建」,Inkling 是其中明牌的一个。\n\n## 美国剩下的「参数缺口」是怎么补上的\n\n报告同一章节给了「分布」视角:把实验室按策略分成两个阵营。\n\n第一阵营是「几乎不发 \u003C 70B 模型」的中国前沿玩家,点名 **Moonshot、Z.ai、Xiaomi、Ant Group**——开发者的第一次接触就是一个跑不起来的超大模型;另一阵营是 Tencent 与 Alibaba Qwen,模型谱覆盖从 \u003C 1B 一直延伸到 2.4T(对应 Qwen3.8-Max 这种旗舰)。报告把第一阵营成立的归因拆成两步:「造大」这件事本身已经不再是差异化能力(Xiaomi、Meituan、Ant Group 都在年内突破万亿参数,而 12 个月前它们在开源圈里还谈不上熟面孔),;同时「社区的量化层」会在几天之内把一个大模型变成本地能跑——「a frontier-first release strategy is viable at all」就靠这条路径撑起来。\n\n美国一边的开源并没有「缺位」,但形态完全不同。报告原文写道:发布新开源模型最多的两家组织,**恰好是做芯片的两家——AMD 与 NVIDIA**,每家年内发布了 200+ 个新模型仓库,远高于第三名 LiquidAI 的约 100 个。报告给出的解读是:硬件厂商意识到开源模型是卖芯片的方式,「一款针对你的硬件优化、又免费开放的模型,是硬件本身能跑的最清晰证明」。在 100B 以上的美国原生模型里,Nemotron 3 Super(124B)、Arcee AI 的 Trinity-Large(399B)是真正从头做的样本,其余多来自对已有模型的转换与「适配」——HF 的原话是「a distribution and optimization layer」。\n\n## 许可证反而指向相反方向\n\n参数天花板的差距之外,报告第三章节《Open weights shift where value accumulates》对许可证做了一次跨地域切片。**178 个中国发布的、参数 > 20B 的模型里,59% 用 Apache 2.0,22% 用 MIT,几乎没有任何模型附带非商用限制**;同一参数带里,美国模型只有 29% 走 Apache 或 MIT,**41% 用自定义条款,30% 干脆没声明许可证**。DeepSeek 和 Z.ai 把 700B–1.65T 级别的模型直接挂在 MIT 下;而美国在同尺寸带的「自定义条款」占比反而最高。\n\n不过 HF 自己也提醒了一句警示:Kimi K3 与 Qwen 3.8 2.4T 最近开始往许可证里加入**非商用限制与营收分成条款**——也就是说,「中国开源比美国更宽松」这个趋势,在最大的那几个模型上正在被悄悄改写。\n\n## 「开源」到底是不是商业模式\n\n报告把这一节的结论说得很直接:**Whatever these releases are for, it is not licence revenue.** 权重以「最宽松的条款免费放出」,回收路径只能在别处——API 与云业务、硬件与平台占位、生态位本身。Z.ai 与 Moonshot 的估值走势被报告当作「开源路线也能走通商业闭环」的注脚;但同一节末尾又给了一句反向提示:**the industry is likely to shift toward clearer monetization paths from open-source adoption**。开源权重正在从「抢占生态」的工具,变成「最终走向商业化」的入口——这与「参数天花板」的故事是同一个故事的两面:中国实验室先把规模抢到手,再把许可证收紧;美国实验室把规模让出去,把硬件层与平台层锁紧。\n\n回到那张月度参数图:2026 年的开源前沿不是一条直线,而是两条几乎不交叉的线。一条从 754B 一路爬到 2.78T,另一条长期贴在 130B 以下,只在两个月被 Nemotron 3 Ultra 与 Inkling 顶上来——而这两个例外里,至少有一个明确承认自己「leverage artifacts from Chinese labs」。开源权重的「地理」正在以参数为度量单位被重新画线,而这条线目前并不在美国那一侧。\n\n参考资料:[Hugging Face — State of Open Models: Summer 2026 Observations](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026)(2026-08-14)。","https:\u002F\u002Fraw.githubusercontent.com\u002Fhuggingface\u002Fblog\u002Fmain\u002Fstate-of-open-models-summer-2026.md","24d5c6c5-6573-4180-a1fd-f1459842d1af",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"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},"0abc35d3-37de-47c0-b378-606d1f341ee5","en","Parameter ceiling or parameter gap: how China set 2.78T as the open-weights frontier in 2026","Hugging Face Summer 2026: In almost every month of 2026, the largest open-weight model from a Chinese lab sat between 754B and 2.78T parameters, while U.S. labs failed to reach 130B in 5 of the 7 months. The only two exceptions — NVIDIA Nemotron 3 Ultra (561B) and Thinking Machines Inkling (952B) — and the report explicitly flags Inkling as \"built on top of Chinese models.\"","## A number hidden by the quantization layer\n\nOn August 14, Hugging Face published \"State of Open Models: Summer 2026 Observations,\" a half-year look at the open-weights ecosystem from January through August 2026. The widely-quoted finding — Qwen hitting 150K derivatives, GGUF repositories up 464% — came from the same report, but the most striking chart lives in the very first section, \"The frontier is moving fast,\" and is titled **Largest open-model releases from Chinese and US labs by month in 2026**. It pins down, in pure numbers, the geopolitical weight of the word \"open.\"\n\n## The real \"parameter ceiling\" gap\n\nThe report states the fact directly: **in almost every month of 2026, the largest and most performant open model from a Chinese lab was larger than any model an American lab released.** Concretely: Chinese frontier labs' monthly ceiling sat between **754B and 2.78 trillion parameters**, while American labs failed to reach 130B in **5 of the 7 months**.\n\nThe two American flagships that did break that pattern are NVIDIA's **Nemotron 3 Ultra** (561B total \u002F 55B active, May and June 2026) and Thinking Machines Lab's **Inkling** (952B). The report names Nemotron 3 Ultra and Inkling together as the only exceptions, then adds a more striking note: **\"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).\"** In other words, a meaningful share of the U.S. open-weights frontier above 100B is \"built on Chinese models,\" and Inkling is the openly acknowledged case.\n\n## How the U.S. filled the parameter gap\n\nThe same section gives a \"strategy\" view, splitting labs into two camps.\n\nCamp one: Chinese frontier players that almost never publish below 70B — **Moonshot, Z.ai, Xiaomi, Ant Group**. A developer's first encounter with them is a model too big to run. Camp two: **Tencent** and **Alibaba Qwen**, covering the whole range from under 1B all the way up to 2.4T (Qwen3.8-Max). The report attributes Camp one's viability to two steps. First, \"building large\" stopped being a differentiator — Xiaomi, Meituan, and Ant Group all cleared a trillion parameters this year, and none were household names in open weights twelve months ago. Second, the community's quantization layer turns a giant model into a locally-runnable one within days — a dependency that makes \"a frontier-first release strategy viable at all.\"\n\nU.S. open source is not absent, but the shape is different. The two organizations publishing the most new open models this year, the report says, are the ones making the hardware: **AMD and NVIDIA**, each with 200+ new model repositories — far ahead of third-place LiquidAI at roughly 100. The interpretation: hardware vendors realized that open models are how you sell chips — \"a model optimized for your hardware and freely available is the clearest proof that the hardware works.\" Among genuinely original U.S. models above 100B, Nemotron 3 Super (124B) and Arcee AI's Trinity-Large (399B) are the true from-scratch samples; the rest comes from converting and adapting existing models — what HF calls \"a distribution and optimization layer.\"\n\n## The licenses point the opposite way\n\nBeyond the parameter ceiling, Section 3, \"Open weights shift where value accumulates,\" cross-cuts by region. Of **178 Chinese releases above 20B parameters**, **59% carry Apache 2.0 and 22% carry MIT, and almost none carry non-commercial restrictions**. In the same size band on the U.S. side, only 29% are Apache or MIT; **41% sit under custom terms, and 30% declare nothing at all**. DeepSeek and Z.ai ship 700B–1.65T models under plain MIT; U.S. labs at the same scale have the highest share of \"custom terms.\"\n\nHF adds one warning sign: **Kimi K3 and Qwen 3.8 2.4T have recently begun adding non-commercial restrictions and revenue-share requirements** to their licenses. The \"China more permissive than the U.S.\" trend is being quietly rewritten at the very top of the curve.\n\n## So what business is \"open\"?\n\nSection 3's conclusion is blunt: **\"Whatever these releases are for, it is not licence revenue.\"** Weights go out under the most permissive terms available; the return has to come from somewhere else — API and cloud, hardware and platform positioning, or the ecosystem position itself. The valuations of Z.ai and Moonshot are cited as proof that \"open\" can close a commercial loop. But the section ends with a reversal: **\"the industry is likely to shift toward clearer monetization paths from open-source adoption.\"** Open weights is moving from a tool for seizing the ecosystem to a funnel that funnels into commercialization. That is the second face of the same story as the parameter ceiling: Chinese labs grab scale first, then tighten licenses; U.S. labs cede scale, then lock down hardware and platform layers.\n\nReturn to the monthly parameter chart: the 2026 open-source frontier is not one line but two that almost never cross. One runs from 754B up to 2.78T; the other stays below 130B and only gets lifted twice — by Nemotron 3 Ultra and by Inkling — and at least one of those two openly admits to \"leveraging artifacts from Chinese labs.\" The geography of open weights is being redrawn in parameters, and the line is not currently on the American side.\n\nReference: [Hugging Face — State of Open Models: Summer 2026 Observations](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026) (2026-08-14).","hf-summer-2026-china-open-weight-parameter-ceiling","2026-08-20T06:00:00Z","2026-08-20T01:07:15.689097Z","2026-08-20T01:07:15.689105Z",true,"agent",74,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"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":46,"title":47,"news_slug":48,"published_at":49},"1844afb1-3a1c-4acd-9e4c-f5e2792a2018","下载免费不等于商用免费：HF Summer 2026 隐藏的开源前沿许可证分水岭","frontier-license-shift-hf-summer-2026","2026-08-23T12:30:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"4bb31ede-b9c4-4762-86ae-9d3b008557ca","Hugging Face Summer 2026 报告:Qwen 拿下 15 万衍生模型, GGUF 仓库一年涨 464%","hugging-face-state-of-open-models-summer-2026","2026-08-18T02:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"5314fe6d-ba17-42bc-9f52-197b8cb9cf91","黄仁勋力挺中国开源大模型:中美技术差距共识正在被开源生态改写","jensen-huang-china-open-source","2026-07-24T03:35:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"c8b1fd9a-524e-4037-adea-d850696291c2","微软测试 DeepSeek V4 接入 Copilot：开源 LLM 首次威胁到头部办公软件的核心","microsoft-copilot-deepseek-v4-open-source","2026-06-22T16:30:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"44d392ff-cace-40f3-a849-14cc7c3893ee","Hugging Face 前高管：开源权重让中国开发者敢把 token「用到极致」","hf-ex-apac-open-weights-china-dev-extreme","2026-06-16T14:00:00+00:00"]