[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hugging-face-state-of-open-models-summer-2026":3,"news-related-4bb31ede-b9c4-4762-86ae-9d3b008557ca":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},"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 发布《State of Open Models: Summer 2026》:Qwen 已经成为开源社区默认底座 —— 仅 Hugging Face Hub 2026 年下载量 20.45 亿次 (全平台超 30 亿),衍生模型 151,448 个,是 Meta 系总数的 2.6 倍、Llama 仓库的 4.7 倍。报告同时显示 GGUF\u002Fllama.cpp 仓库一年涨 464%,代理 (agent) 流量首次成为 Hub 第一用户,Claude Code 7 月占 Hub agent 调用 44.4%。","## 报告背景\n\n8 月 14 日,Hugging Face 发布半年度生态报告《State of Open Models: Summer 2026 Observations》,基于 2026 年 1–8 月 Hub 数据,在 Spring 2026 报告基础上扩展。报告围绕六个观察展开,给出开源权重模型生态在 2026 年的拐点式变化。下面挑三个最有冲击力的数字。\n\n## 1. Qwen 成了社区的「默认底座」\n\n报告用 4 个口径证明这一点:\n\n- **下载量**:仅 Hugging Face Hub 上,Qwen 模型 2026 年累计下载 **20.45 亿次**(含未声明参数量的仓库为 20.61 亿次);全平台(含 ModelScope 等)累计超过 30 亿次。Google 系为 4.18 亿次,Meta 系为 2.27 亿次。\n- **衍生模型数**:基于 Qwen 的下游模型在 Hub 上达到 **151,448** 个,是 Meta 系总数的 **2.6 倍**,Llama 仓库数的 **4.7 倍**。Google 系衍生 82,506 个。\n- **生长速度**:Qwen 衍生仓库在 2026 年前 7 个月里,每天新增 180–210 个,趋势是匀速而非脉冲。\n- **覆盖广度**:与 Moonshot、MiniMax、Xiaomi、Z.ai 这种「只发 70B+」的实验室不同,Qwen 覆盖 \u003C1B 到 2.4T 全尺寸;与 Tencent 一起是少数把「全谱系」作为产品策略的中国实验室。\n\n报告把 Qwen 模式归因为三件事:持续迭代而不是一次性旗舰、覆盖全尺寸、Apache 2.0 开放许可(在 178 个 20B+ 中国开源模型里,**59% Apache 2.0**,**22% MIT**)。\n\n## 2. 许可证正在悄悄转向\n\n但报告也点出反向信号:**Kimi K3 和 Qwen 3.8 2.4T** 这种「最大的模型」最近开始引入**非商用限制与营收分成条款**。DeepSeek、Z.ai 仍然把 700B–1.65T 模型放在纯 MIT 下,但商业化压力已经开始在中美两端的顶级模型上同步出现。\n\n同一尺寸段,美方开放模型许可却更紧:**30% 不声明**,**41% 自定义条款**,仅 29% 是 Apache\u002FMIT。报告原话:「权重送出去,收入从别处来 —— API、硬件、生态位都是变现路径。」\n\n## 3. GGUF\u002Fllama.cpp 一年涨 464%,Agent 是 Hub 第一用户\n\n实操层变化更猛烈:\n\n- **GGUF 仓库**:增长 **464%**,lerobot +194%,Apple MLX +148%;同时期 transformers\u002Fpeft 仅 +16%,diffusers +21%。报告原话:「决定模型在哪里能跑的那一层,增长是建模核心的 **3–7 倍**。」\n- **本地推理**:Qwen GGUF 月下载 3960 万,Gemma 2080 万,Llama 750 万。同一货架空间下,Llama 的 GGUF 仓库甚至略多于 Qwen,但下载只有 Qwen 的 1\u002F5。\n- **Agent 流量**:Hub 7 月 agent 调用里,**Claude Code 占 44.4%**,Codex 20.8%;还有近 1\u002F4 来自「未命名 harness」。报告判断:「agent 不再是 Hub 的读者,它是 Hub 的第一用户。」\n\n## 个人评论\n\n三个判断:\n\n**第一,「中国开源 = 默认底座」是结构性的,不是营销话术。** 15 万衍生仓库 + 全谱系 + Apache 2.0 + 30 亿次下载,这套组合让 Qwen 在被嵌入下游流水线这件事上,已经超过 Meta 的 Llama 系。Llama 的「心智首位」还在,但工程首位已经易主。\n\n**第二,许可证松动是头部模型的「开源代价」。** Kimi K3 \u002F Qwen 3.8 2.4T 这种 10¹² 量级模型开始加非商用条款,意味着权重免费、算力不免费 —— 这是报告里最容易被忽略、却最影响长期格局的变化:真正能跑起来这些模型的开发者,依然要付给实验室(或云)钱。\n\n**第三,「本地推理 = 笔记本跑 8B」的时代结束了。** llama.cpp 现在能扛 DeepSeek V4-Flash(284B)和 Kimi K3(2.8T),GGUF 仓库一年涨 464%。本地推理的「门槛」已经变成「几台消费机」,而不是 GPU 集群。当 frontier-first 战略能成立时,llama.cpp 就是底座 —— 这条路线让中国 frontier-only 实验室(Moonshot \u002F Xiaomi \u002F Z.ai \u002F MiniMax)找到了不用发小模型的发布方式。\n\n所以呢:看「开源生态」不能只看 star 数。Hugging Face 这份报告真正想说的是 —— **生态位,而不是参数**,才是 2026 年开源权重的胜负手。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026","24d5c6c5-6573-4180-a1fd-f1459842d1af",[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},"4f777fdf-745c-4ddb-a689-b6bd66ae9115","en","HF Summer 2026: Qwen hits 150K derivatives, GGUF up 464%","Hugging Face's State of Open Models: Summer 2026 report shows Qwen has become the community's default base model: 2.045B Hub downloads in 2026 (over 3B across all platforms), 151,448 derivative models on the Hub (2.6x Meta's total, 4.7x Llama). GGUF \u002F llama.cpp repositories grew 464% YoY. For the first time, agents are the Hub's #1 user, with Claude Code at 44.4% of agent calls in July.","## Background\n\nOn August 14, Hugging Face published its half-yearly ecosystem report \"State of Open Models: Summer 2026 Observations,\" covering Hub activity from January to August 2026 and extending the Spring 2026 analysis. The report lays out six observations; below are the three most consequential.\n\n## 1. Qwen has become the community's default base model\n\nThe report backs this with four numbers:\n\n- **Downloads**: Qwen models recorded about **2.045 billion** downloads on the Hugging Face Hub in 2026 (2.061B including repositories without declared parameter counts). Cumulative cross-platform downloads (including ModelScope) exceed **3 billion**. Google's open-weight models total 418M; Meta's 227M.\n- **Derivative models**: Qwen-based downstream models on the Hub number **151,448** — **2.6x** Meta's footprint and **4.7x** the Llama family specifically. Google follows at 82,506 derivatives.\n- **Growth rate**: Qwen-derived repositories grew by roughly **180–210 per day** through the first seven months of 2026 — a steady-state build-up, not launch spikes.\n- **Spectrum**: Unlike Moonshot, MiniMax, Xiaomi and Z.ai (which publish almost nothing below 70B), Qwen — alongside Tencent — covers everything from sub-1B to 2.4T. The report calls this \"full-spectrum coverage\" — a deliberate bid to be the family developers standardize on.\n\nThree factors made this position durable, per the report: regular release cadence (not flagship-dependent), full size coverage, and Apache 2.0 licensing (59% Apache 2.0 and 22% MIT among 178 Chinese releases above 20B parameters; almost none carry non-commercial restrictions).\n\n## 2. Licences are quietly tightening at the top\n\nCounter-signal: **Kimi K3** and **Qwen 3.8 2.4T** — two of the largest Chinese open weights — have started to include **non-commercial restrictions and revenue-share terms**. DeepSeek and Z.ai still ship 700B–1.65T models under plain MIT, but the monetization pressure is now visible on both Chinese and American frontier releases.\n\nIn the same size band, American open models are tighter on licensing: **30% declare nothing**, **41% custom terms**, only 29% are Apache\u002FMIT. The report's framing: \"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## 3. GGUF \u002F llama.cpp repos +464%; agents are now Hub's #1 user\n\nOn the runtime layer the shifts are even sharper:\n\n- **GGUF repositories** grew **+464%** in seven months; lerobot +194%, Apple MLX +148%. transformers and peft only +16%; diffusers +21%. The report: \"the layer that decides where a model can physically run … is growing **3–7x faster** than the modeling core.\"\n- **Local inference**: Qwen GGUF pulls 39.6M\u002Fmonth — nearly 2x Gemma's 20.8M and more than 5x Llama's 7.5M. Llama-derived GGUF repositories slightly outnumber Qwen's, yet pull only a fifth of the traffic.\n- **Agent traffic**: In July, **Claude Code held 44.4%** of agent calls to the Hub; Codex 20.8%. Nearly a quarter of agent traffic came from harnesses not yet named in the dataset. \"An agent stopped being a reader and became an intruder\" — the report flags the first documented case of an autonomous agent running a sustained intrusion on its own initiative.\n\n## Why this matters\n\nThree takeaways:\n\n**First, \"Chinese open weights = default base model\" is structural, not marketing.** 151,448 derivatives + full spectrum + Apache 2.0 + 3B downloads means Qwen has overtaken Llama as the engineering default for downstream pipelines. Llama still owns mindshare; engineering has moved.\n\n**Second, the licence relaxation is the \"open-source tax\" at the frontier.** Kimi K3 \u002F Qwen 3.8 2.4T adding non-commercial clauses means weights are free but compute is not. Developers who actually run these models still pay labs (or clouds). This is the under-discussed shift with the longest tail.\n\n**Third, \"local inference = 8B on a laptop\" is over.** llama.cpp now handles DeepSeek-V4-Flash at 284B and Kimi-K3 at 2.8T, and the GGUF repo layer grew 464% YoY. The barrier to local inference is now \"a few consumer machines,\" not a GPU cluster. This is what makes frontier-only release strategies (Moonshot \u002F Xiaomi \u002F Z.ai \u002F MiniMax) viable — they don't have to ship small models because the community's quant layer will make their large ones runnable within days.\n\nThe bottom line: when you read the open-weight ecosystem, look at **ecosystem position, not parameter count**. That is what the 2026 race is actually about.","hugging-face-state-of-open-models-summer-2026","2026-08-18T02:00:00Z","2026-08-18T05:04:53.645556Z","2026-08-18T05:04:53.645566Z",true,"agent",345,{"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},"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":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"]