[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mistral-small-4-firefox-smart-window":3,"topics-all":38,"news-related-cb1e799d-d6d7-4ab9-9eaf-bea0aa432b06":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},"cb1e799d-d6d7-4ab9-9eaf-bea0aa432b06","Mistral 模型进驻 Firefox:119B 开放权重模型驱动浏览器 AI 助手","Mozilla 与 Mistral 宣布合作:Firefox AI 浏览助手 Smart Window beta 改由 Mistral Small 4 驱动,总参 1190 亿、激活约 65 亿,法国与北美先行;对话默认不存 Mozilla 服务器,合作方零数据保留。","9 月 16 日,Mozilla 与法国 AI 公司 Mistral 联合宣布:Firefox 的 AI 浏览助手 Smart Window(仍处 beta 阶段)即日起改由 Mistral 模型驱动,先面向法国和北美用户,英国与德国预计年内跟进。对 Mistral 而言这是一次少见的转身——这家公司过去主要服务企业客户,如今借浏览器直接触达全球消费者。\n\n## 哪个模型在驱动\n\n据 Techzine 报道,Smart Window 背后是 Mistral Small 4:总参数 1190 亿的开放权重模型,每次推理只激活约 65 亿参数。这个体量显然无法塞进消费级设备本地运行,但少量激活参数让推理成本维持在合理区间——对一个要向全体浏览器用户开放的助手来说,这笔账必须算得过来。功能上,Smart Window 做三件事:梳理复杂搜索、记住浏览过但已关闭的重要信息、根据当前标签页查找关键信息。\n\n## 隐私架构是这次合作的卖点\n\nMozilla 给 Smart Window 配了一套偏保守的隐私框架:功能完全 opt-in,只使用用户主动选择共享的标签页与历史记录;对话内容默认不存储在 Mozilla 服务器上,Mistral 等合作伙伴承诺零数据保留;网页检索与来源展示则交由另一家强调隐私的 AI 公司 Exa 完成。这套设计延续了 Firefox 的传统定位,也是它敢把 AI 深度集成进浏览器的底气。\n\n## 为什么是 Mistral\n\nMistral 在[官方公告](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-x-mozilla\u002F)中给出四个理由,核心一条是\"开放技术需要开放分发\":Mozilla 为开放 Web 战斗了二十多年,Mistral 从第一款模型起就坚持开放权重,两家希望借此证明开源 AI 也能服务全球消费者。Mistral 还强调其模型在区域语言、方言和文化语境上做过微调,目标是让各地用户得到\"母语感\"而非\"翻译腔\"的回答。Mozilla Corporation CEO Anthony Enzor-DeMeo 说得更直白:浏览器不该是\"单向漏斗\",而应保留互联网自由探索的底色。Mistral 联合创始人兼 CEO Arthur Mensch 则把这概括为共同推进隐私、控制与选择权。\n\n## 别急着乐观\n\nTechzine 的评论泼了冷水:Small 4 是约半年前发布的模型,以当下标准看并不惊艳,DeepSeek-V4.1 Flash 在 Artificial Analysis 这类智能基准上明显领先,代价是参数量约为其五倍。这次合作更像\"够用的模型配上对的分发渠道\",而非模型能力的胜利。另一边,浏览器阵营对 AI 本就分裂:Vivaldi 承诺永不集成 AI,联合创始人 von Tetzchner 的口号是\"keep browsing human\"。Firefox 在隐私老用户与 AI 新功能之间走钢丝,结果要交给市场检验。\n\n对行业来说,这件事的信号价值大于产品本身:开放权重模型的竞争,正在从 benchmark 榜单延伸到分发渠道。当浏览器这个使用时长最高的入口之一开始向非巨头模型开放,下一轮模型之争的胜负手,可能不在实验室的评测集里,而在每台设备的默认设置里。","https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-x-mozilla\u002F","2436174c-644b-4a65-9a98-e7a3b705569a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"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},"348fd89e-6dcc-41d3-a805-30d49064b95e","en","Mistral Small 4 now powers Firefox Smart Window","Firefox's Smart Window beta now runs on Mistral Small 4 (119B params, ~6.5B active). France and North America first; opt-in, zero data retention.","On September 16, Mozilla and French AI company Mistral jointly announced that Firefox Smart Window, the browser's AI assistant (still in beta), is now powered by Mistral models. The rollout starts in France and North America, with the United Kingdom and Germany expected to follow later this year. For Mistral it marks an unusual turn: a company that has traditionally served enterprises now reaches consumers directly through a browser.\n\n## Which model is doing the work\n\nAccording to Techzine, Smart Window runs on Mistral Small 4, an open-weight model with 119 billion total parameters and only about 6.5 billion activated per inference. That is nowhere near small enough to run locally on consumer hardware, but the small number of activated parameters keeps inference cost in a reasonable range — a calculation that matters for an assistant opened up to all browser users. Functionally, Smart Window does three things: makes sense of complex searches, remembers important pages you clicked away from, and finds key information based on your current tabs.\n\n## Privacy architecture is the selling point\n\nMozilla wrapped the feature in a deliberately conservative privacy frame. Smart Window is fully opt-in and only uses tabs and history that users choose to share. Conversations are not stored on Mozilla's servers by default, and partners like Mistral agree to zero data retention. Web retrieval and source display are handled by Exa, another company that brands itself privacy-respecting.\n\n## Why Mistral\n\nMistral's [official announcement](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmistral-x-mozilla\u002F) gives four reasons, with \"open technology needs open distribution\" at the core: Mozilla has spent more than two decades fighting for an open web, and Mistral has released open-weight models since its first release — together they want to prove open-source AI can serve consumers worldwide. Mistral also stresses that its models are fine-tuned on regional languages, dialects and cultural context, so answers feel native rather than translated. Mozilla Corporation CEO Anthony Enzor-DeMeo put it more bluntly: a browser shouldn't be a \"one-way funnel\", and should preserve the internet's original freedom to explore. Mistral co-founder and CEO Arthur Mensch frames the deal as jointly advancing privacy, control and choice.\n\n## Don't get carried away\n\nTechzine's editorial is skeptical: Small 4 was released about half a year ago and is unremarkable by current standards; DeepSeek-V4.1 Flash is far ahead on intelligence benchmarks like those from Artificial Analysis, at the cost of roughly five times the parameters. The partnership looks more like \"an adequate model meeting the right distribution channel\" than a victory of model capability. Meanwhile the browser camp is split on AI: Vivaldi pledges never to integrate it, with co-founder von Tetzchner's slogan \"keep browsing human\". Firefox is walking a tightrope between privacy-minded old users and new AI features, and the market will judge.\n\nFor the industry, the signal matters more than the product itself: competition among open-weight models is extending from benchmark leaderboards to distribution channels. When the browser — one of the highest-engagement entry points — opens up to non-giant models, the next round of the model race may be decided not in lab eval sets, but in the default settings of every device.","mistral-small-4-firefox-smart-window","2026-09-16T17:07:00Z","2026-09-16T17:08:28.088469Z","2026-09-16T17:08:28.088488Z",true,"agent",7,[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},"5535b4e4-21de-4ed6-a9bb-2b4d824e6568","F-Droid 一次更新的 102 款应用,72.5% 主要是 AI 写的","f-droid-72-percent-ai-written-audit","2026-09-15T15:12:16+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"dcd8b3e1-a3c7-4614-aba4-9002219ea5f6","LibreDB Studio 0.15 发布:本地 LLM 接管数据库交互","libredb-studio-local-llm-agent","2026-09-15T00:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"58ed753e-ad6d-4aac-95f4-36bf217e169c","把 10 万条人类视频变成机器人教材:RoboTok 检索 mAP 提升约 50 倍,hard 任务 79.3% 对 19.5%","robotok-retrieval-benchmark-reread","2026-09-06T21:11:25+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"005557c5-8a3c-4d34-89bc-35d5351c4570","蒸馏只需要一条训练样本?清华实测:单条query覆盖71.5%训练状态,16条追平17k全量","one-shot-opd-single-query-distillation","2026-09-05T21:07:11+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"7623f190-7071-4811-a6f1-32462a99b8d3","经验会过期:阿里云论文让自主后训练的有害授权率从 62.5% 降到 25%","bcit-conditional-experience-transfer-post-training","2026-09-05T17:11:11+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"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"]