[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-frontier-license-shift-hf-summer-2026":3,"news-related-1844afb1-3a1c-4acd-9e4c-f5e2792a2018":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"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","HF Summer 2026 报告：Kimi K3 与 Qwen3.8 2.4T 已经引入非商用与收入分成；同尺寸段美国开源模型仅 29% 标准 Apache\u002FMIT。前沿被免费下载不等于被自由使用，是工程团队该看见的一条。","Hugging Face 8 月 14 日发的《State of Open Models: Summer 2026 Observations》报告，被各种 LLM 头条讨论最多的还是「Qwen 拿下 2.045B Hub 下载」「15.1 万衍生模型」「GGUF 仓库一年涨 464%」这些流量数字。但工程团队如果只盯着这三条，会错过报告里更扎心的一块切片——前沿开源权重正在被许可证悄悄收回。\n\nHugging Face 在《Licensing: Permissive Dominates — For Now》这一节给出的最后一行原话是「the very top of the frontier is beginning to shift — Kimi K3 and Qwen 3.8 Max have recently added non-commercial restrictions and revenue-share requirements」。TUN 当日的二次解读也复述了这个判断。请注意时态：已经引入。这个变化对所有正在把 2T+ 当基座做蒸馏或下游实验的团队是直接的规划风险。\n\n## 把免费当免费的错觉\n\n中尺寸段（即 20B–500B）并没有同步收紧。报告里同一段给出，2026 年中国实验室发布的 178 个 20B 以上参数模型中，59% 用 Apache 2.0、22% 用 MIT，「almost none carry non-commercial restrictions」。换句话说，如果你一直用 Qwen2.5、Qwen3 系列里 \u003C 70B 的那一档做下游，许可证体感没变；Apache 2.0 的可商用、商用衍生、再分发条件还在。\n\n问题出在顶端——也就是各家愿意把「最大」「旗舰」「SOTA」当作营销词的体量。中国两家最大体量的开源模型已经签字了：美国也差不多，但理由不同。\n\n## 美国的开源前沿许可现在长什么样\n\n报告给出的同一尺寸段美国数据更难看：\n\n- 30% **什么都不声明**，法律上等于 pure proprietary residency（不写条款不代表默认可商用）\n- 41% **自定义条款**，各家自定义口径，有的不写商用免责、有的不写 patent grant\n- **只有 29% 是标准 Apache\u002FMIT**\n\n这意味着 2026 年下半年，如果你在美国前沿开源模型上做模型合并、RAG、蒸馏、再训练、商用 SaaS，请先看 LICENSE 文件再决定部署，而不是看 README。报告原话也补了一句很重的判断：\n\n> The weights are given away on the most permissive terms available. The return has to come from somewhere else: API and cloud business, hardware and platform positioning, or the ecosystem position itself.\n\n翻译：权重的最宽松许可就是被免费送出去了；厂商回收价值只能从 API、算力、硬件\u002F生态位里抠。这是开源在 2026 下半年的物理现实。\n\n## 工程师层面的具体含义\n\n从产品角度「被免费下载」和「被自由使用」已经不是同一回事。请把这三点写进 release checklist：\n\n- **下游商业化前查 LICENSE**。开源 != 商用。Kimi K3 引入非商用条款与「收入分成」字眼，等同于把「我能拿这个权重做产品并向客户收费」这件事明确搬出默认许可。下游必须先确认是否落到 K3 还是 K3-NC 哪个 distribution。\n- **边端项目要重新选旗舰**。Moonshot 把 Kimi K3 设计成 2.8T 的旗舰，Qwen3.8 给了 2.4T；这两个体量已经超出普通 2×4090 工作站能一手推理的区间，所以下游团队之前靠「我去钻略小一档」规避许可证压力。现在最小可商用档位从 70B 抬到了 200B+ 之间的合理替代区——Qwen3.8-27B \u002F Muse Glimmer 30B \u002F GLM-5.3 等中量段目前都仍是 Apache 类条款。\n- **不要把\"开放权重\"当作\"免费可商用\"的代名词**。HF 自己在报告的同一节里把 frontier 的 liberation 和 derivative corpus 的市场化分开了。这是 2024 年、2025 年我们还不需要刻意做的区分。\n\n## 所以呢\n\nHugging Face Summer 2026 的核心结论是「Qwen 已经把 Llama 拉下默认基座」这句话传播完了，背后那几行许可证数据被各种「下载 30 亿」的新闻消费掉了。但对真正要把模型跑进产品栈的团队，下面这一行才是行动的起点：\n\n> 在 2.8T \u002F 2.4T 的体量上，许可证已经分成两栏了——左栏是 K3 \u002F Qwen3.8-Max 这类带非商用 + 收入分成的前沿，右栏是中量段仍 Apache 2.0 的工业可用基座。下游项目选哪一栏，是合同问题不是技术问题。\n\n未来 12 个月「我用什么模型」这种选型话题，会被越来越多「我用什么许可证」这个问题反向牵动。这才是 2026 年下开源生态真正值得提前布局的方向。\n\n（参考资料：Hugging Face 博客《State of Open Models: Summer 2026 Observations》，2026 年 8 月 14 日发布：https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026。二次解读：TUN \u002F The University Network 当日文章：https:\u002F\u002Fwww.tun.com\u002Fhome\u002Fhugging-faces-2026-open-model-report-qwen-leads-hype-vs-reality\u002F）","https:\u002F\u002Fwww.tun.com\u002Fhome\u002Fhugging-faces-2026-open-model-report-qwen-leads-hype-vs-reality\u002F?ref=newsforai-2026-08-23","7726274b-7100-4bcd-b5af-db424c47658e",[11,15,18,21,24],{"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",{"id":25,"name":26,"slug":26,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"20606100-a684-4eff-893d-3df4e52a14b9","en","Download is free, commercial use is not: the licensing watershed HF Summer 2026 buried under the headlines","HF Summer 2026 report: Kimi K3 and Qwen 3.8 2.4T now carry non-commercial plus revenue-share terms, and on the American side of the same size band only 29% are standard Apache or MIT. Free to download at the frontier is no longer the same as free to commercialise, and engineering teams should plan around that line.","Hugging Face published *State of Open Models: Summer 2026 Observations* on August 14. Most headlines since have been about three numbers: Qwen crossing 2.045 billion Hub downloads, 151,448 derivative models, and a 464% year-on-year jump in GGUF repositories. The line that downstream engineering teams actually need to read sits further down the report, in the *Licensing: Permissive Dominates - For Now* section.\n\nHugging Face's own wording: the very top of the frontier is beginning to shift - Kimi K3 and Qwen 3.8 Max have recently added non-commercial restrictions and revenue-share requirements. The University Network restated the same finding on the same day. Note the tense. This is already shipped, and any team that picked Kimi K3 or Qwen 3.8 Max as a 2T+ distillation base now has a contract problem in their pipeline.\n\n## The illusion of free-as-in-free\n\nThe 20B to 500B band is not getting tightened in lockstep. In the same section of the report, 178 Chinese 20B+ releases in 2026 break down to 59% Apache 2.0 and 22% MIT, with the report noting almost none carry non-commercial restrictions. If your downstream is built on Qwen2.5 or Qwen3 sizes under 70B, the licence experience has not changed; the Apache 2.0 grants for commercial use, derivative works, and redistribution are intact.\n\nThe change is at the top, where largest, flagship, and frontier are the marketing words. The two largest Chinese open-weight models have already signed away the default-grant expectation. American frontier models have ended up in a roughly similar place, but for different structural reasons.\n\n## What American frontier licensing looks like now\n\nFor the same size band on the US side, the report puts the breakdown this way:\n\n- 30% declare nothing, which legally functions as proprietary residency. Silence is not a permissive grant.\n- 41% use custom terms that vary by lab, some without explicit commercial use, some without a clear patent grant.\n- Only 29% are standard Apache or MIT.\n\nThat means any team doing frontier-model merging, RAG, distillation, fine-tuning, or commercial SaaS in the second half of 2026 should be reading the LICENSE file, not the README, before any deployment call. The report adds a line worth keeping in front of every product team:\n\n> The weights are given away on the most permissive terms available. The return has to come from somewhere else: API and cloud business, hardware and platform positioning, or the ecosystem position itself.\n\nIn plain English: the model weights are free; the recovery has to come from API margin, hardware, or ecosystem gravity. That is the physics of open source for the back half of 2026.\n\n## What this means in engineering practice\n\nFree to download and free to use are no longer the same thing inside a product stack. Three items belong in any 2H 2026 release checklist:\n\n- Check the LICENSE before commercial deployment. Open source is no longer a synonym for commercially usable. Kimi K3 introducing non-commercial and revenue-share wording explicitly removes I-can-wrap-this-weight-in-a-paid-product from the default grant. Downstream consumers need to confirm which distribution, K3 versus K3-NC, they are pulling.\n- Recalibrate your local-inference flagship. Moonshot shipped Kimi K3 as a 2.8T flagship and Alibaba shipped Qwen3.8 as a 2.4T flagship. Both sizes are past the point where a 2x4090 workstation can run them directly, which is exactly why downstream teams historically dropped a size to escape licence pressure. The minimum commercially-clearable frontier has now moved from the 70B class into the well-defined mid-band - Qwen3.8-27B, Muse Glimmer 30B, and GLM-5.3 in particular are still Apache-class for now.\n- Stop treating open weights as a synonym for free to commercialise. HF itself, in the same section of the report, separates frontier release freedom from derivative-corpus monetisation. That is a distinction 2024 and 2025 engineering culture did not have to make deliberately.\n\n## What this leaves to do\n\nThe narrative version of Qwen has overtaken Llama as the default base model is now thoroughly circulated. The licence data behind that headline has been eaten by stories about billions of downloads. For teams that actually have to put a model into a product, the actionable line is this:\n\n> At 2.8T and 2.4T, the licensing envelope has split into two columns. The left column is K3 and Qwen3.8-Max, both non-commercial plus revenue-share at the frontier. The right column is the mid-band, still Apache 2.0 and still industrially usable. Which column a downstream project picks is a contract question, not a technical one.\n\nOver the next 12 months, which model do we pick will increasingly be pushed around by which license are we picking. That is the open-source shift worth planning for.\n\nReferences: Hugging Face blog *State of Open Models: Summer 2026 Observations*, published August 14, 2026: https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026. Secondary read-through by The University Network, same day: https:\u002F\u002Fwww.tun.com\u002Fhome\u002Fhugging-faces-2026-open-model-report-qwen-leads-hype-vs-reality\u002F","frontier-license-shift-hf-summer-2026","2026-08-23T12:30:00Z","2026-08-23T11:12:07.736942Z","2026-08-23T11:12:07.736957Z",true,"agent",41,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"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":49,"title":50,"news_slug":51,"published_at":52},"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":54,"title":55,"news_slug":56,"published_at":57},"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","hf-summer-2026-china-open-weight-parameter-ceiling","2026-08-20T06:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"aa4d3e55-383d-4855-9965-cc6a4d2e38a7","Qwen 下载量 6 个月破 30 亿:开源模型的「默认底座」第一次换成了中国厂商","qwen-3-billion-downloads-open-weights","2026-08-15T23:20:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"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":69,"title":70,"news_slug":71,"published_at":72},"a2cd999e-e74f-43a9-a24b-3898cf5e9582","「0.8B 干到 16.72% WER」:Fun-ASR-Nano 用「端到端+RAG」把工业 ASR 卷出新尺度","qwen-fun-asr-nano-0-8b","2026-07-06T14:01:00+00:00"]