[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nvidia-acquires-hugging-face-open-source-ai":3,"topics-all":38,"news-related-7bae3d71-a5c2-4588-95e7-b5d4b5c7085a":57},{"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},"7bae3d71-a5c2-4588-95e7-b5d4b5c7085a","开源模型 4.4 个月追上闭源前沿:Hugging Face 被 NVIDIA 129 亿美元收编","Mozilla 9 月发布的《State of Open Source AI》v1.1 披露,英伟达 9 月 3 日以 129.3 亿美元收购开源 AI 平台 Hugging Face。1800 万开发者使用的模型与数据集枢纽,正式并入 NVIDIA 资产负债表。","9 月 3 日,Jensen Huang 在 NVIDIA 官方博客宣布以 129.3 亿美元收购 Hugging Face。这笔交易让拥有 1800 万开发者、托管超过 300 万个模型和 50 万个数据集的全球最大开源 AI 仓库,正式进入算力霸主版图。这也是 NVIDIA 历史上第二大并购,仅次于去年底 200 亿美元收购 Groq 资产。\n\nHugging Face CEO Clément Delangue 向 CNBC 透露,这次收购由 Hugging Face 主动发起——他们在今年夏天意识到开源 AI 到了转折点,需要更多资源、规模与可见度,而 NVIDIA 是最合适的归宿。Huang 在博文中承诺 Hugging Face 仍是开放的平台,开发者可以自由选择模型、框架、云和硬件,NVIDIA 算力不会是强制要求。\n\n## 把这次收购放进 Mozilla 9 月报告里看\n\n把这条交易放进 Mozilla 同日发布的《State of Open Source AI》v1.1 报告里看,意义远超一次普通的科技并购。Mozilla 的数据显示,2026 年 8 月,DeepSeek 旗下开源模型历史上首次超越 Google,在 OpenRouter 上登顶请求量榜首;当月全球请求量前 10 名模型中,有 8 个是开放权重模型,其中 7 个由中国团队打造。同期,NVIDIA 自家也已向 Hugging Face 贡献了超过 500 个模型和 250 个开源数据集,是平台最大的开源贡献方。\n\n报告还给出了一个耐人寻味的资本版图:整个开源 AI 栈已经被买空了。NVIDIA 收购了 Hugging Face,Stripe 以约 75 亿美元收购 OpenRouter,Cohere 与 Aleph Alpha 合并估值约 200 亿美元,CoreWeave 拿下 Weights & Biases,Databricks 收编 MosaicML,NVIDIA 还接连拿下了 Run:ai、Gretel、OctoAI。算力、模型市场、MLOps、推理优化——每一层都落入了某个大厂的资产负债表。\n\n## 开源追到只差 4.4 个月\n\n但 Mozilla 同时指出,开源模型能力追上闭源前沿的速度比多数人想象的快。报告依据 METR 时间视野数据拟合得出:开放权重模型与闭源前沿的差距大约是 4.4 个月,Epoch AI 给出的估计是 4 个月左右。Artificial Analysis 智能指数上,Kimi K3 与 GLM-5.3 同分 60,排在 Claude Opus 5(63)、Claude Fable 5(62)、GPT-5.6 Sol(61)、Grok 4.6(61)之后,与第一梯队只差两三分。\n\n价格上的反差更直接。GLM-5.3 标价每百万 token 1.15\u002F3.50 美元,而 Claude Fable 5 是 10\u002F50 美元——同档能力下,价格差到 5–10 倍。Mozilla 用 Terminal-Bench 2.1 数据测算,在统一 harness 下,GLM-5.2 每任务 0.43 美元,Claude Opus 4.8 是 2.41 美元;同档模型能力,开放权重便宜 5 倍多。\n\n## 真正稀缺的是入口,不是模型\n\n把这两条线放在一起看,这次收购反映的是开源 AI 的某种成熟悖论:开放权重模型跑得越来越快、卖得越来越便宜,真正的稀缺资源反而落到了「分发入口」「算力调度」「训练 loop」这三层。NVIDIA 把 1800 万开发者活跃的入口吃下,意味着开源生态的「门面」被算力控制方收编。\n\n黄仁勋在博文中强调开源环境给防御者带来「非对称优势」——防守方比攻击方多,透明协作能放大这一差距。Mozilla 在报告中引用了一个具体案例:今年 7 月那次 Hugging Face 安全事件中,OpenAI 拒绝用前沿模型读取攻击日志,被 Hugging Face 用 GLM-5.2 自托管重建了取证管道,挖出比纯文本扫描多 4 倍的凭据。\n\n所以这次收购对开源 AI 究竟是好是坏,答案取决于 NVIDIA 是否真的让 Hugging Face 保持中立。如果它继续做多云、多硬件、多框架的公共仓库,开源生态只是换了个有钱的房东;如果它把流量导向自家推理栈,GPT-5 那种「闭源 + 唯一入口 + 锁定生态」的模式会在开源侧重演一次。\n\nMozilla 的报告最后一句话写得很克制:开放权重的开发者应该被请进决策 AI 的那些房间里,与闭源厂商平等坐在一起。Hugging Face 卖给 NVIDIA 之后,这间屋子又少了一张椅子。","https:\u002F\u002Fblogs.nvidia.com\u002Fblog\u002Fnvidia-to-acquire-hugging-face\u002F","474eef8c-e0c3-46cf-adee-c089558220f9",[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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",{"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},"44c8fef5-6503-45b6-9957-9f36764587a9","en","Open models close the gap to 4.4 months as NVIDIA buys Hugging Face for 2.9B","Mozilla's State of Open Source AI v1.1 report released in September 2026 reveals that NVIDIA agreed on September 3 to acquire open-source AI platform Hugging Face for 2.93 billion. The hub used by 18 million developers now sits on the chip giant's balance sheet.","On September 3, Jensen Huang announced on NVIDIA's official blog that the company had agreed to acquire Hugging Face for $12.93 billion. The deal hands the world's largest open-source AI hub — 18 million developers, more than 3 million models and 500,000 datasets — into the balance sheet of the chip giant. It is NVIDIA's second-largest acquisition on record, behind the $20 billion Groq assets purchase at the end of last year.\n\nHugging Face CEO Clément Delangue told CNBC that the acquisition was initiated by Hugging Face itself. Over the summer, the company realized that open-source AI had reached a turning point and needed more resources, scale and visibility; NVIDIA was the natural home. Huang committed in the post that Hugging Face will remain an open platform — developers can choose models, frameworks, clouds and hardware; NVIDIA compute will not be required.\n\n## Why this deal matters beyond M&A\n\nRead alongside Mozilla's State of Open Source AI v1.1 report, released the same day, the transaction means more than a routine tech acquisition. Mozilla's data shows that in August 2026, a DeepSeek open model overtook Google for the first time to take the top spot for requests on OpenRouter. Eight of the top ten models by request volume that month were open-weight; seven of those eight were Chinese-built. NVIDIA itself has contributed more than 500 models and 250 open datasets to Hugging Face, making it the platform's largest open contributor.\n\nThe report also lays out a striking capital map: the entire open-source AI stack has been bought up. NVIDIA bought Hugging Face; Stripe agreed to buy OpenRouter for roughly $7.5 billion; Cohere and Aleph Alpha merged at an enterprise value around $20 billion; CoreWeave acquired Weights & Biases; Databricks took over MosaicML; NVIDIA also snapped up Run:ai, Gretel and OctoAI. Compute, model routing, MLOps and inference optimization — every layer now lives on someone's corporate balance sheet.\n\n## Open is closing the gap to 4.4 months\n\nMozilla also points out that open models are catching up to the closed frontier faster than most people think. The report fits METR time-horizon data and finds a roughly 4.4-month gap between open-weight models and the closed frontier; Epoch AI estimates around four months. On the Artificial Analysis Intelligence Index, Kimi K3 and GLM-5.3 tie at 60, sitting behind Claude Opus 5 (63), Claude Fable 5 (62), GPT-5.6 Sol (61) and Grok 4.6 (61) — only two to three points off the leader group.\n\nThe price gap is even sharper. GLM-5.3 lists at $1.15 \u002F $3.50 per million tokens, versus Claude Fable 5 at $10 \u002F $50 — a 5–10x spread at the same capability tier. On Terminal-Bench 2.1 under a neutral harness, Mozilla's data shows GLM-5.2 at $0.43 per task and Claude Opus 4.8 at $2.41; same capability, more than 5x cheaper on the open side.\n\n## The real scarcity is the front door, not the model\n\nTaken together, the acquisition reflects a mature paradox in open-source AI. Open-weight models run faster and cheaper every quarter; what becomes scarce instead is the distribution front door, compute scheduling, and the training loop. NVIDIA taking over the front door that 18 million developers actively use means the open ecosystem's storefront has been absorbed by the compute incumbent.\n\nHuang emphasizes in his post that open environments give defenders an asymmetric advantage: defenders outnumber attackers, and transparent collaboration amplifies that gap. The Mozilla report cites a concrete case — in the July Hugging Face security incident, OpenAI refused to use frontier models to read attack logs, so Hugging Face rebuilt the forensics pipeline on self-hosted GLM-5.2 and recovered roughly four times the credentials a plain-text scan had found.\n\nWhether this deal is good or bad for open-source AI therefore comes down to whether NVIDIA genuinely keeps Hugging Face neutral. If it stays multi-cloud, multi-accelerator and multi-framework, the open ecosystem just got a richer landlord. If it routes traffic toward its own inference stack, the GPT-style pattern — closed weights, single entry point, locked ecosystem — will repeat itself on the open side.\n\nMozilla's report closes with a restrained line: developers who keep AI open, portable, and widely deployed should be seated in the rooms where AI gets decided, on equal footing with closed providers. After Hugging Face went to NVIDIA, that room has one fewer chair.","nvidia-acquires-hugging-face-open-source-ai","2026-09-17T08:00:00Z","2026-09-17T09:04:57.725649Z","2026-09-17T09:04:57.725657Z",true,"agent",80,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"367476a4-b9af-46f1-a6ab-3de1d83640ff","NVIDIA 把中国开发者日搬到苏州:10 月连开两天,AI 推理和物理 AI 是主菜","nvidia-china-developer-day-2026-suzhou","2026-09-16T03:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"54b86d93-0fd0-4107-9353-9b79a1446f69","NVIDIA 开源 IMO 金牌完整配方:30\u002F42 分、561B 双专家、算力账本全公开","nvidia-nemotron-imo-gold-open-recipe","2026-09-11T17:13:27+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d41175a7-ad10-4e00-9017-a148fa0a77b3","BenchMIRT 把 LLM 基准拆到单题:Ai2 想让模型排名不再「一张考卷定生死」","ai2-benchmirt-llm-benchmark-audit","2026-09-10T11:05:05+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"61cf8d85-c751-4da2-9aae-10b645415ec9","英伟达发布开源工具 PAIR,把家里电脑连成个人 AI 推理集群","nvidia-pair-personal-ai-router-local-inference","2026-09-09T02:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"42b7939c-1b44-43b8-95cf-a8fc2204560d","NVIDIA 开源 Personal AI Router，把家里 RTX 与 Mac 拼成本地 AI 集群","nvidia-personal-ai-router-pair-beta","2026-09-04T03:20:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"f2f0017f-449d-492b-b9d7-90a2013023fb","英伟达 129 亿美元收购 Hugging Face 接近敲定:开源 AI 仓库终被算力霸主收编","nvidia-12-9-billion-hugging-face-acquisition","2026-09-04T03:00:00+00:00"]