[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hugging-face-summer-2026-attention-adoption":3,"news-related-1311adb6-dc19-41a7-a188-6760d9e53672":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},"1311adb6-dc19-41a7-a188-6760d9e53672","HF Summer 2026 报告:13 个下载量 Top 25 模型是 2022 年的老面孔","Hugging Face 2026 夏季报告:Hub 下载 Top 25 模型中有 13 个是 2022 年的,与点赞榜几乎不重合;all-MiniLM-L6-v2 7 个月下载 15.5 亿次,仅 5,156 likes。中国前沿实验室下载集中于 ≥70B 模型,美实验室基本无大模型被下载。","2026 年 8 月 14 日, Hugging Face 发布《State of Open Models: Summer 2026 Observations》。在头部厂商们追逐更大、更新的模型时, 这份报告用 Hub 数据反问了一件很多人忽略的事情: 被点赞最多的, 和真正在被跑的, 几乎不是同一批模型。\n\n## 数据本身就很反直觉\n\n报告取了 2026 年 Hub 上下载量 Top 25 的模型仓库, 与点赞 Top 25 的仓库对比。**两个榜单重合的只有一个**。在 2026 年新发布的所有模型里, 没有一个进入下载榜 Top 25; 而下载榜里有 13 个模型来自 2022 年。\n\n最具代表性的数字是 all-MiniLM-L6-v2 —— 一个 2022 年发布的 sentence embedding 模型。它在 7 个月内被下载了约 15.5 亿次, 同时只积累了 5,156 个 likes。同一时期, Kimi-K3 每个 like 对应约 60 次 pull。一个被业界忽略, 一个被业界仰望。\n\n## \"点赞\"和\"下载\"记录的是两件不同的事\n\n报告点出了这件事的本质: **likes 是注意力, downloads 是依赖**。一个 like 说的是\"这个发布值得关注\", 主要流向发售后几周内的前沿模型; 一次 download 说的是\"这个模型被接进了某条按计划运行的 pipeline\", 主要流向运行稳定的小型模型。**把任何一个指标当成另一个指标的代理, 是观察 Hub 时最常见的错误** —— 报告原文就这么写。\n\n数据分布的不对称同样出现在发布者层面:\n\n- **中国 frontier 实验室** 的下载几乎都集中在 70B 以上参数。MiniMax 2026 年的下载里 100% 来自 ≥70B 模型; Moonshot 88%; DeepSeek 55%; Z.ai 39%。Moonshot 这样的 frontier-only portfolio 全年累计 37M 下载。\n- **美国实验室** 这一栏基本是空的。Google、Microsoft、IBM Granite 在 ≥70B 区段几乎没有下载; NVIDIA 14%, Meta 仅 9%。\n- 总量上, Qwen 的 full-spectrum 策略覆盖从亚 1B 到万亿参数, 全年达到 2,045M 下载 —— **是 Moonshot 整年下载量的约 55 倍**。\n\n报告强调 Qwen 累计拥有 151,448 个衍生模型, 是 Meta 全部衍生模型数量的 2.6 倍、Llama 衍生模型数量的 4.7 倍, 并且仍在以每天 180-210 个新仓库的速度增长。\n\n## 另一条主线: 模型之外的事长得更快\n\n报告专门用一节讨论 runtime layer, 因为真正决定一个模型能不能被用上的, 往往不是模型本身, 而是它跑在哪。\n\n- 仓库声明 `gguf` 库(供 llama.cpp 使用)同比增长 464%\n- `lerobot` 增长 194%\n- 苹果 MLX 增长 148%\n- 而 `transformers`、`peft` 这类核心建模库只增长 16%\n- `diffusers` 21%\n\n报告提到的另一件事是 **ggml 团队 2026 年 2 月正式加入 Hugging Face**。项目保持完全开源、社区治理, 但获得了更稳定的资源支持。本地推理的\"上限\"由此被抬到了一个新的层级: GGUF 版本现在能装下 2.8 万亿参数的 Kimi-K3 与 284B 的 DeepSeek-V4-Flash。\n\n这也解释了一个看似矛盾的事实: 几乎所有顶尖实验室都不怎么发布官方 GGUF 版本, 但开发者本地跑模型时用的恰恰是 GGUF。Qwen 每月被以 GGUF 形式下载 3,960 万次, 是 Gemma(2,080 万)的近两倍、Llama(750 万)的五倍多。\n\n## 一个还没被写出来的判断\n\n这件事最值得关注的点, 是它改变了我们判断\"开源模型成功\"的尺子。在 frontier 模型榜单上拼排名, 看 likes 也许够用; 但谁真正被接进了生产 pipeline, 看 downloads 才有意义。报告里的描述值得记住: \"Likes are the right instrument for reading what the field is excited about, downloads for reading what it currently depends on.\"\n\n过去一年, 评价一个开源模型的方式几乎都是\"哪个最 fancy\"。当 2026 年有 60% 以上的衍生仓库都指向 Qwen 家族, 当 15.5 亿次下载流向一个 sentence embedding 老模型, 当 GGUF 仓库一年涨 464%, \"被用上\"和\"被注意\"是两件事就不再是一个抽象观察, 而是一个可以直接拿来修正报道框架的判断。\n\n做下游产品的人, 把 model card 上的 likes 数当成生产可用性的替代指标, 这个习惯大概需要重新校准一次。","https:\u002F\u002Fwww.plushcap.com\u002Fcontent\u002Fhuggingface\u002Fblog\u002Fhuggingface-state-of-open-models-summer-2026-observations","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},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"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},"47b93abe-96f1-4974-bd0b-e85be374dd0c","en","HF Summer 2026 Report: 13 of the Top 25 Most-Downloaded Models Are from 2022","Hugging Face's Summer 2026 report: 13 of the Hub's top 25 most-downloaded models are from 2022, and the list barely matches the top 25 by likes. all-MiniLM-L6-v2 was downloaded 1.55 billion times in seven months against just 5,156 likes. Chinese frontier labs concentrate downloads in ≥70B models, while U.S. labs see essentially no downloads of large models.","Hugging Face published \"State of Open Models: Summer 2026 Observations\" on August 14, 2026. While frontier labs race to ship the largest, newest weights, the report quietly reframes a question the field has been measuring the wrong way: which models people are excited about, and which models are actually running in production, are not the same list.\n\n## The data itself is counter-intuitive\n\nThe authors took the top 25 model repositories by downloads accumulated over 2026 and the top 25 by likes. **Exactly one repository appears in both lists.** No model published in 2026 reaches the download top 25, while thirteen of the twenty-five date from 2022.\n\nThe most striking number is for `all-MiniLM-L6-v2`, a sentence embedding model released in 2022. It was pulled roughly 1.55 billion times in seven months, against only 5,156 likes. In the same window, Kimi-K3 was pulled about 60 times per like. One is invisible to the hype cycle; the other dominates it.\n\n## Likes and downloads record different acts\n\nThe report draws a clean line between the two metrics: **likes is attention, downloads is dependency**. A like says \"this release matters\" and accrues to flagship releases within weeks of launch; a download says \"this model is wired into a pipeline that runs on a schedule\" and accrues to small, stable models over years. The authors note that using either as a proxy for the other is the most common mistake in coverage of the Hub.\n\nThe asymmetry shows up at the publisher level too:\n\n- **Chinese frontier labs** get essentially all their 2026 downloads from models above 70B parameters. 100% of MiniMax's 2026 downloads are above 70B, alongside 88% of Moonshot's, 55% of DeepSeek's, and 39% of Z.ai's. Moonshot's frontier-only portfolio recorded 37M downloads over the year.\n- **American labs** look different. Google, Microsoft and IBM Granite record essentially none of their 2026 downloads above 70B; NVIDIA only 14%, Meta only 9%.\n- By total downloads, Qwen's full-spectrum strategy reached 2,045M downloads in 2026 — roughly **55 times Moonshot's frontier-only total**.\n\nThe report highlights Qwen's 151,448 derivative models on the Hub, 2.6× Meta's total footprint and 4.7× Llama derivatives specifically, growing at 180-210 new repositories per day.\n\n## The other headline: things outside the model are growing faster\n\nThe report devotes a section to the runtime layer, because what actually decides whether a model can be used is usually not the model itself, but where it runs.\n\n- Repositories declaring the `gguf` library (used by llama.cpp) rose **464%** year over year\n- `lerobot` rose 194%\n- Apple MLX 148%\n- Meanwhile core libraries like `transformers` and `peft` grew only 16%\n- `diffusers` 21%\n\nThe other piece is that **the ggml team joined Hugging Face in February 2026**. The project remains fully open-source and community-governed, but now has durable resources behind it. The local inference ceiling moved with it: GGUF builds now carry DeepSeek-V4-Flash at roughly 284B parameters and Kimi-K3 at roughly 2.8 trillion.\n\nThis explains a contradiction that looks bizarre on the surface: the labs behind the ten largest model families publish very few official GGUF conversions, yet developers running models locally overwhelmingly use GGUF. Qwen gets 39.6 million GGUF downloads per month, nearly twice Gemma's 20.8 million and more than five times Llama's 7.5 million.\n\n## The takeaway the field hasn't fully internalized\n\nThe most useful sentence in the report is also its most understated: \"Likes are the right instrument for reading what the field is excited about, downloads for reading what it currently depends on.\"\n\nFor a year, the open-source model conversation has been measured almost entirely on which release was the most \"fancy.\" When 60% of new derivative repositories point at the Qwen family, when a 2022 sentence embedding model gets pulled 1.55 billion times, when GGUF repositories grow 464%, \"being used\" and \"being noticed\" stop being abstract observations and become something you can use to recalibrate how you report on the field.\n\nFor teams building downstream products, treating a model's like count as a proxy for production readiness is a habit that probably needs another calibration cycle.","hugging-face-summer-2026-attention-adoption","2026-08-24T08:00:00Z","2026-08-24T07:05:17.069288Z","2026-08-24T07:05:17.069298Z",true,"agent",44,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"7b26392a-4bb3-448f-881a-335ad8aee605","QwQ-32B：强化学习驱动的开源推理模型，以320亿参数比肩6710亿的DeepSeek-R1","qwq-32b-rl-rival-deepseek-r1-agent","2026-05-23T14:08:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"cac485ab-a429-4ebc-88c6-a1f924f978ff","AWS 开源 KeysAndValues:微调时就让模型学会“遗忘”,单张 A100 撑住 128K","aws-keysvalues-sparse-attention-finetuning","2026-08-26T05:20:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"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":56,"title":57,"news_slug":58,"published_at":59},"c94bdf86-5de9-49fe-8c98-0f5c47611bfe","SGLang v0.5.18 发布:大模型冷启动提速 2.38 倍,710 个 PR 都改了什么","sglang-v0-5-18-cold-start-2-38x","2026-08-24T23:15:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"988bbfb8-672a-4c6c-98f7-3a170b6bd8b3","Macaw 把 LFM2.5 装进 1.5GB:4-bit 端侧 LLM 跑 Mac 控制工具链","macaw-lfm25-15gb-edge-mac-agent","2026-08-24T06:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"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"]