[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hf-open-models-summer-likes-vs-downloads":3,"news-related-89804e7d-cee8-4412-8b30-e43855911ef5":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},"89804e7d-cee8-4412-8b30-e43855911ef5","点赞与下载是两个经济体：Hugging Face 夏季报告拆穿开源模型的「追新幻觉」","Hugging Face 8 月 14 日发布的夏季开源模型报告里，最反直觉的发现是：2026 年下载量 Top 25 与点赞 Top 25 交集只有 1 个仓库，下载榜没有一个 2026 年新模型、13 个来自 2022 年——点赞衡量兴奋度、下载衡量依赖度，两者几乎不相关。报告同时显示 Moonshot 只发大模型的路线年下载约 3,700 万次，只有 Qwen 全尺寸策略 20.45 亿次的约 1\u002F55，而 Kimi K3 与 Qwen 3.8 2.4T 最近开始在许可证中加入非商业限制和收入分成条款。","## 一份「打脸」榜单\n\nHugging Face 8 月 14 日发布的《State of Open Models: Summer 2026》里藏着一个所有追新的人都不想看到的数据：把 2026 年内累计下载量 Top 25 和点赞数 Top 25 拉出来对，**两份榜单的交集只有 1 个仓库**。更扎心的是，下载 Top 25 里没有一个模型是 2026 年发布的，其中 13 个要追溯到 2022 年。\n\n也就是说，你在时间线上刷到的每一次重磅开源发布，几乎都不在生产环境的依赖清单里。（[原始报告](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026)，本报道基于 HF 官方博客及其 GitHub 镜像）\n\n## 点赞和下载，记录的是两件不同的事\n\n报告作者把话说得很直白：点赞说明一个发布「重要」，集中在新模型发布后的几周；下载说明一个东西「被接进了定时跑的流水线」，经年累月积累在小型、稳定的模型上。原文的判断是——把其中一个当另一个的代理指标，是外界（包括 HF 自己早期的文章）解读 Hub 数据时最常见的错误。\n\n极端样本：老牌句向量模型 all-MiniLM-L6-v2 七个月被拉取 15.5 亿次，点赞只有 5,156 个；话题度最高的 Kimi-K3，大约每 1 个赞对应 60 次下载。两个数字根本不在同一个经济系统里。底层的分布形状也极端：约 85.6% 的模型仓库终身下载量不足 200 次，1.5% 的仓库拿走了 99.2% 的下载。\n\n## 「只发大模型」和「全尺寸铺开」差了 55 倍\n\n按尺寸拆分下载来源，是这份报告里最锋利的一块。MiniMax 2026 年的下载几乎全部来自 70B 以上模型，Moonshot 为 88%，DeepSeek 55%，Z.ai 39%；而 Google、微软、IBM Granite 在 70B+ 档位的下载几乎为零，NVIDIA 和 Meta 也只占 14% 和 9%。\n\n但两条路线的回报差距巨大：Moonshot 的 frontier-only 组合全年录得约 3,700 万次下载，而覆盖从 2.4T 的 Qwen 3.8 Max 到 27B 小模型的全家桶策略，给 Qwen 带来约 20.45 亿次——差了约 55 倍。报告认为尺寸策略已经变成「意图声明」而非「能力声明」：只发前沿是在赌榜单位置和 API 需求，全尺寸铺开是在竞标「开发者标准化到哪个家族」。frontier-only 之所以还能成立，靠的是 llama.cpp 把本地推理的天花板顶了上去——GGUF 构建的 DeepSeek-V4-Flash 约 284B、Kimi-K3 约 2.8 万亿参数，万亿级 MoE 摊到几台消费级机器上跑，在一年前并不存在。\n\n## 最宽松的许可证时代，可能正在见顶\n\n178 个 20B 以上的中国模型里，59% 用 Apache 2.0、22% 用 MIT，几乎都没有非商业限制；DeepSeek 和 Z.ai 甚至在 7000 亿到 1.65 万亿参数的体量上直接用纯 MIT。对照美国同体量区间：只有 29% 是 Apache 或 MIT，41% 是自定义条款，30% 干脆什么都没声明。但就在最近几周，风向开始转——Kimi K3 和 Qwen 3.8 2.4T 这两个最大的模型，开始在许可证里加入非商业限制和收入分成条款。报告直言这些权重不是许可生意，回报要从 API 云业务、硬件生态位或生态位本身来。\n\n## Agent 是新用户，而且还没有赢家\n\n7 月发布的 agent-usage 数据集第一次让「谁在调 Hub」可见。Claude Code 7 月占 agent 流量 44.4%，但单月数据会骗人——它 4 月曾占 67.8%，5 月 64%，而 Codex 从 10.4% 一路爬到 20.8%；7 月还有近四分之一的 agent 流量来自数据集没登记名字的 harness，5 月这个比例高达 59.8%。这是一个没有在位者的市场，一次版本更新就能在一个月里搬走一半流量。HF 在报告结尾直接写：Agent 第一次成为 HF Hub 上排名第一的用户。\n\n## 所以呢\n\n对开发者，这份报告的价值是选型方法论的修正：榜单刷到的新模型衡量的是注意力，你真正要依赖的东西要看下载曲线和衍生生态——追新和铺生产管线是两条不同的采购路径。对行业，注意许可证那一节：当最大的两个开源模型开始谈收入分成，「最宽松」这三个字的保质期可能只剩几个季度。\n\n热度会骗人，依赖不会。","https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fblog\u002Fblob\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},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",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},"5745411e-5328-402e-9e86-fede5d63fd37","en","Likes and downloads are different economies (HF report)","The most counterintuitive finding in Hugging Face's Summer 2026 State of Open Models report: the top 25 repositories by 2026 downloads and the top 25 by likes overlap in exactly one repository, and no model published in 2026 made the download list — thirteen date from 2022. Likes measure excitement while downloads measure dependency, and the two barely correlate. The report also shows Moonshot's frontier-only route recorded roughly 37 million annual downloads — about 1\u002F55 of Qwen's 2,045 million under a full-spectrum strategy — while Kimi K3 and Qwen 3.8 2.4T recently began adding non-commercial restrictions and revenue-share clauses to their licenses.","## A List That Hurts\n\nBuried in Hugging Face's \"State of Open Models: Summer 2026\" report, published August 14, is a number nobody chasing shiny new releases wants to see: cross-reference the top 25 repositories by downloads accumulated in 2026 against the top 25 by likes, and **exactly one repository appears on both lists**. Worse, not a single model published in 2026 made the download top 25 — thirteen of them date back to 2022.\n\nIn other words, almost none of the blockbuster open-weight launches flooding your timeline are in production dependency lists. ([Original report](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fstate-of-open-models-summer-2026); this article is based on the official HF blog post and its GitHub mirror.)\n\n## Likes and downloads record two different acts\n\nThe report's authors put it plainly: a like says a release matters, and concentrates on frontier models in the weeks after they ship; a download says something is wired into a pipeline that runs on a schedule, accruing to small, stable models over years. Their verdict — treating either as a proxy for the other is the most common mistake in coverage of the Hub, including HF's own earlier work.\n\nThe extreme sample: the veteran sentence-embedding model all-MiniLM-L6-v2 was pulled 1.55 billion times in seven months against just 5,156 likes; Kimi-K3, the most talked-about release, saw roughly 60 downloads per like it received. The two numbers live in different economies. The underlying distribution is just as extreme: about 85.6% of model repositories have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads.\n\n## Frontier-only vs. full-spectrum: a 55× gap\n\nSplitting downloads by model size is the sharpest China-US contrast in the report. Effectively all of MiniMax's 2026 downloads came from models above 70B parameters, along with 88% of Moonshot's, 55% of DeepSeek's and 39% of Z.ai's. No large American account looks like this: Google, Microsoft and IBM Granite recorded essentially none of their 2026 downloads above 70B, with NVIDIA and Meta at just 14% and 9%.\n\nThe payoff gap between the two routes is enormous: Moonshot's frontier-only portfolio recorded about 37 million downloads over the year, while Qwen's full-spectrum strategy — spanning the 2.4T-parameter Qwen 3.8 Max down to 27B variants — reached about 2,045 million. A factor of roughly 55. The report argues a size profile is now a statement of intent rather than capability: a frontier-only portfolio bets everything on benchmark position and API demand; a full-spectrum portfolio is a bid to be the family developers standardize on. Frontier-only remains viable at all because llama.cpp pushed the local-inference ceiling upward — GGUF builds of DeepSeek-V4-Flash at roughly 284B and Kimi-K3 at roughly 2.8 trillion parameters mean a trillion-parameter mixture-of-experts spread across a few consumer machines, something that simply did not exist a year ago.\n\n## The most permissive licensing era may be peaking\n\nOf 178 Chinese releases above 20B parameters this year, 59% carry Apache 2.0 and 22% carry MIT, almost none with non-commercial restrictions; DeepSeek and Z.ai ship models between 700 billion and 1.65 trillion parameters under plain MIT. The American side of the same size band: only 29% Apache or MIT, 41% under custom terms, and 30% declaring nothing at all. But in the last few weeks the wind shifted — Kimi K3 and Qwen 3.8 2.4T, the two largest models, started adding non-commercial restrictions and revenue share requirements to their licenses. The report is blunt: whatever these releases are for, it is not license revenue; the return has to come from API and cloud business, hardware and platform positioning, or the ecosystem position itself.\n\n## Agents are the new user — with no incumbent\n\nThe agent-usage dataset published in July made \"who is calling the Hub\" visible for the first time. Claude Code led July with 44.4% of agent traffic, but a single month conceals the real finding — it held 67.8% in April and 64% in May, while Codex climbed steadily from 10.4% to 20.8%. Nearly a quarter of July's agent traffic came from harnesses not yet named in the dataset; in May that figure was 59.8%. This is a market with no incumbent, where one release or one changed default can move half the traffic in a month. HF closes the report stating it directly: agents became the number-one user of the HF Hub for the first time.\n\n## So what\n\nFor developers, the report's value is a correction to selection methodology: the new models flooding your feed measure attention; what you actually depend on is read from download curves and derivative ecosystems — chasing novelty and provisioning production pipelines are two different procurement paths. For the industry, watch the licensing section: when the two largest open models start talking about revenue share, the shelf life of \"most permissive\" may only be a few quarters.\n\nHype deceives. Dependencies don't.","hf-open-models-summer-likes-vs-downloads","2026-08-18T13:20:00Z","2026-08-18T13:12:24.025124Z","2026-08-18T13:12:24.025136Z",true,"agent",100,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"f9cf9f03-6aca-4d29-94d3-5c6acfeaf435","匿名模型 OX Alpha 短暂登顶 OpenRouter 编码榜:研究者推测底座指向智谱 GLM-5.x","ox-alpha-stealth-openrouter-glm-5-zhipu","2026-08-24T03:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"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-summer-2026-frontier-ceiling","2026-08-22T14:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"22a1a718-0eb6-46e5-8ee8-825400de11d1","DeepMind WeatherNext 在 Nature 发论文：用 28 km 粗分辨率做出多一天的飓风预警,代码权重全部开源","deepmind-weathernext-cyclones-nature-open-source","2026-08-10T02:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"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":66,"title":67,"news_slug":68,"published_at":69},"ed8ef087-9dc6-4295-bbdd-d0d2c92977d3","GCC 拒绝 LLM 生成的实质性贡献:开源基础设施开始为 AI 代码划红线","gcc-rejects-llm-contributions-15-line-threshold","2026-07-30T03:30:00+00:00"]