[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mistral-mozilla-firefox-smart-window-moe":3,"topics-all":38,"news-related-4497a0c5-e9b8-42d5-8d72-cfa0a49c1fba":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},"4497a0c5-e9b8-42d5-8d72-cfa0a49c1fba","Mistral Small 4 加入 Firefox Smart Window：开放权重模型第一次进浏览器助手默认菜单","9 月 16 日,Mistral Small 4 加入 Firefox Smart Window 模型选择器,法国首次开通,北美、加拿大已上线。Apache 2.0 的 119B\u002F6.5B 激活 MoE 是四个预设里出 token 单价最贵的,Mozilla 5.31% 桌面份额把 Mistral 送到几亿台机器前。","9 月 16 日,Mistral 与 Mozilla 同时发文,宣布 Mistral 模型成为 Smart Window(Firefox 内置浏览 AI 助手)的可选引擎,法国本月的首个新闻点,北美与加拿大是下一个主要市场;Mistral 的两位创始人把这次合作包装成\"开源技术 + 开源分发\"的路线图。表面看像是又一笔浏览器内置 AI 合作,但仔细读两份通稿会发现一个微妙的落差:Mozilla 把 Mistral Small 4 描述成\"新增的一个模型选项\",Mistral 那边则用\"已渲染好\"的口径,通稿措辞差异背后是开放权重模型第一次坐进浏览器助手默认菜单这件事本身。\n\n## 不是排行榜上的,是把模型送进浏览器分发\n\nFirefox Smart Window 早在 8 月 18 日就上线了,这次合作的核心变化是 Smart Window 原本有三个预设选项——Fast 用的是 Google 的 Gemini Flash Lite,Flexible 用的是阿里 Qwen3-235B-A22B-Instruct-2507,Personal 用的是 OpenAI 的 gpt-oss-120b——Mistral Small 4 加入后成为第四个预设,但 Mozilla 在表意里特地保留了一句\"多家 AI 模型\"的措辞,意味着 Mistral 并没有把其它三家挤掉,只是新增了一个位。与此同时,Mistral 的表意用\"is now powered by\"(\"现由 Mistral 模型驱动\"),明显比 Mozilla 通稿更具排他意味,但这份话术里没有点 Mistral Small 4 的名字,这是 Mistral 母版与 Mozilla 之间一次明显存在\"翻译损耗\"的握手。\n\nMistral 选在这时候落子也是深思熟虑。8 天前的 9 月 8 日,Mistral 完成了一轮 30 亿欧元的融资,投后估值约 210 亿欧元(约 240 亿美元),三星领投,EQT 与 PSG Equity 联合跟投。一年前 ASML 领投的 C 轮估值是 117 亿欧元,十二个月涨了约 79%。Mistral 之前的商业模式走的是企业与公共部门,从来没有面向消费者的分发通路,这次合作正好补上了 Mozilla 端着 5.31% 全球桌面份额这个数字(8 月 StatCounter),Chrome 的 73.28% 是它的 13.8 倍。所以 Mozilla 给 Mistral 一个入场位,Mistral 给 Mozilla 一个 Chrome 没法抄的故事,这是典型的双向背书而不是纯技术选型。\n\n## 开放权重里最贵的预设:Mistral Small 4 不是\"小\"\n\nMistral Small 4 是这次合作的实际模型选型,3 月 16 日发布,Apache 2.0 协议,API 名是 mistral-small-2603。最容易让人误读的一点是它的\"小\"字——119B 总参数,但每个 token 只激活 6.5B,典型 MoE 设计:128 个专家,每次推理选 4 个走通。这种设计上的\"宽解释\"在 Mistral 的官方口径里被描述成\"用大模型的能力付小模型的推理成本\"。上下文窗口 262,144 tokens,支持文本 + 图像输入,Apache 2.0 协议意味着 Mozilla 可以照宣传话术那样称之为\"开源合作\"而不是\"供应商集成\"。\n\n放在 Smart Window 的预设队列里看,Mistral Small 4 是单 token 激活参数二轻的,gpt-oss-120b 激活 5.1B \u002F 总 116.8B,Mistral Small 4 激活 6.5B \u002F 总 119B,Qwen3-235B-A22B 激活 22B \u002F 总 235B,Google 没公布 Gemini Flash Lite 的参数量;但出 token 价格层面,Mistral Small 4 在四个里最贵,每百万出 token 的挂牌价 0.60 美元,gpt-oss-120b 0.17 美元,Qwen3-235B-A22B 0.35 美元,Gemini 2.5 Flash Lite 0.40 美元。Mozilla 自己掏钱为每一个 Smart Window 会话结算,这意味着尽管 Mistral 在名单里标着\"开放权重\",它放在 Mozilla 的成本表上反而是最贵的那个——选开放权重的开源标签不必然等于选便宜。\n\n## 隐私口径的字面读法\n\n两份通稿都把隐私放在最显眼的位置。Mistral 那边的原话是:\"对话默认不保存在 Mozilla 的服务器上,Mistral 等合作伙伴承诺零数据留存。\"这两句话的主语不一样——前者是 Mozilla 的基础设施,后者是 Mistral 的基础设施。Mozilla 那边的\"默认\"是一个配置项,配置可被改;Mistral 的\"零数据留存\"是 API 层的承诺,意味着不持久化 prompt 与 completion,不用于训练。两者结合,才是 Smart Window 在读你打开的标签页时真正能依赖的承诺。\n\n但有一个问题被两份通稿都绕过去了:实际推理发生在哪个数据中心。Mistral 那边反复用\"主权 AI\"(\"sovereign AI\")这种语汇词,适合代表欧洲场景下本地化部署的实际诉求,但 Mozilla 与 Mistral 都没在通稿里点出具体地域,如果有数据驻留义务的买家需要在 AI Controls 里自己核实。\n\n## 四个预设的实质意义\n\n打开 Smart Window 的 AI 模型选择器,目前是四件套加一个 Custom 位——四件套是 Fast(Gemini)、Flexible(Qwen)、Personal(gpt-oss-120b)和新加入的 Mistral Small 4,Custom 位允许指向自己的端点,Mozilla 在支持文档里专门提示本地模型在某些能力上不一定稳定。三件套是开放权重,只有 Gemini Flash Lite 是闭源;四件套全部以托管服务的形式存在,选 Apache 2.0 模型不等于在你的本地跑推理,只等于跑在别人机器上的权重是公开可审查的——这是两种不同的承诺,但不是同一种保障。\n\n这件事放在 2026 年的浏览器助手版图里看,意义不止 Mozilla 多了一个模型选型。Vivaldi 已经公开承诺不集成助手,把\"保持人类式浏览\"作为它的入口形式;Mozilla 选了反例。SideCar 这边,Mistral 显然押的是另一条相反方向——当 Mozilla 是中立分发者,选 Mistral 时,模型能力被欧洲本地化拉通,把\"主权 AI\"和\"开源权重\"打通,以一种 Mozilla 标签里的\"四件套\"形式落到用户身上。Claude 的 Anthropic、OpenAI 的 GPT、Qwen 的模型都坐在同一架菜单里,这一件事本身就是对开放权重在消费级场景的最可能落点的一种标定。\n\n那么对于作为每天使用开源模型的开发者来说,这件事说明 K 个层面:开放权重模型不再只是企业 API 里的备选项,它开始坐到消费级助手入口的菜单里;Apache 2.0 的权重与 0.60 美元的出 token 单价并存的现实意味着开源不等于便宜,选模型时仍然要看每 token 的实际账单;浏览器作为分发入口的话语权比模型本身的领先程度更值得关注——Mistral 模型比 Anthropic 落后多少还在可争议范围,但 Mozilla 5.31% 的桌面份额能直接把 Mistral 推到几亿台机器前,这件事已经是既成事实。","https:\u002F\u002Fblog.mozilla.org\u002Fen\u002Ffirefox\u002Fmozilla-mistral-partnership\u002F","2436174c-644b-4a65-9a98-e7a3b705569a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",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},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"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},"a3b796f7-372e-49a1-b7c2-ac875f081dd1","en","Mistral Small 4 enters Firefox Smart Window: open weights reach the browser assistant menu","On 16 September, Mistral Small 4 was added to the Firefox Smart Window AI model picker. France opened for the first time, with North America and Canada already live. The Apache 2.0 119B\u002F6.5B-active MoE model is the most expensive preset per output token; Mozilla's 5.31% desktop share puts the French lab in front of several hundred million machines.","On 16 September 2026, Mistral and Mozilla posted back-to-back announcements: Mistral models are powering Firefox Smart Window, the browser's in-built AI browsing assistant, with France as the first new market and North America and Canada already live. Read side by side, the two posts tell subtly different versions of the same deal, and that mismatch is the most interesting part of the story. Mozilla describes Mistral Small 4 as \"a new AI model\" added to the picker; Mistral's newsroom says the feature \"is now powered by Mistral models\". The same day, two different verbs. It is also the first time a French open-weight model has sat at the front of a browser assistant menu with an Apache 2.0 license attached.\n\n## Not the leaderboard — putting the model into browser distribution\n\nFirefox Smart Window launched on 18 August 2026 and was updated on 16 September to add the Mistral integration and the France rollout. The capability list did not change. Before this deal, the model picker exposed three presets plus a Custom endpoint: Fast ran Google Gemini Flash Lite, Flexible ran Alibaba's Qwen3-235B-A22B-Instruct-2507 and Personal ran OpenAI's gpt-oss-120b. Mistral Small 4 becomes the fourth preset, sitting alongside rather than displacing the others.\n\nThe choice of timing is not accidental. Eight days earlier, on 8 September 2026, Mistral closed a €3 billion round at a post-money valuation north of €21 billion, led by Samsung Electronics and co-led by EQT's Scaleup Europe Fund and PSG Equity. A year earlier, Mistral's C-round had been led by ASML at a €11.7 billion post, putting the twelve-month increase at roughly 79%. Mistral's business has historically been enterprise and public-sector — the company had never had a consumer distribution channel. Mozilla, meanwhile, sits at 5.31% of worldwide desktop browsing in August 2026, behind Chrome on 73.28%, Edge on 10.46% and just ahead of Safari on 5.24% (StatCounter). Chrome is roughly 13.8 times Firefox's desktop share. Mozilla is not distributing Smart Window from a position of strength, and Mistral is not arriving from one either; the deal is mutual reinforcement, not a takeover of either side.\n\n## The most expensive preset — Mistral Small 4 is not actually small\n\nMistral Small 4 shipped on 16 March 2026 under Apache 2.0 (API name mistral-small-2603). The \"small\" in the name is misleading on its own — the model carries 119B total parameters but activates only 6.5B per token, a sparse mixture-of-experts design with 128 experts and 4 active at any moment. The 256k-token context window accepts text and image input. Mistral's own material claims a 40% reduction in end-to-end completion time and 3× requests per second versus its predecessor; those are vendor figures rather than independent benchmarks and should be read as such.\n\nInside the Smart Window picker, Mistral Small 4 sits as the second-lightest per token. gpt-oss-120b activates 5.1B of 116.8B, Mistral Small 4 activates 6.5B of 119B and Qwen3-235B-A22B activates 22B of 235B; Google does not publish parameter counts for Flash Lite. On list pricing per million output tokens, however, Mistral Small 4 is the most expensive preset at $0.60, against $0.17 for gpt-oss-120b, $0.35 for Qwen3-235B-A22B and $0.40 for Gemini 2.5 Flash Lite. Mozilla pays the bill on every Smart Window session, which means picking the open-weight label does not necessarily mean picking the cheap option.\n\n## Reading the privacy wording literally\n\nBoth companies lead with privacy. Mistral's exact words: \"conversations aren't saved on Mozilla's servers by default, and partners like Mistral agree to zero data retention.\" Two sentences, two different subjects. The first is about Mozilla's infrastructure and carries a \"by default\" qualifier; the second is about Mistral's infrastructure and carries no qualifier at all. Mozilla's wording explicitly notes a control surface exists in AI Controls — the out-of-the-box configuration does not retain, but a setting can change.\n\nZero data retention is a standard enterprise API term: the provider does not persist prompts or completions beyond the request and does not train on them. It is a stronger claim than \"we don't sell your data\", and it is the commitment that actually matters for a Smart Window session reading your open tabs.\n\nNeither post names a datacenter region. Mistral's post uses \"sovereign AI\" framing throughout, but the inference location is not on the record; any buyer with a data-residency obligation needs to verify that themselves.\n\n## What the four-preset lineup actually means\n\nThe Smart Window picker is a four-preset plus Custom lineup, with three of the four presets open-weight and only Gemini Flash Lite closed. All four still run as hosted services — selecting an Apache 2.0 model inside Smart Window does not mean inference happens on the user's machine. It means the weights running on someone else's machine are publicly inspectable. That is a meaningfully different guarantee, but it is not the same as local inference.\n\nThe lineup matters beyond Mozilla picking one more model. Vivaldi has publicly pledged against browser-level AI integration, framing the goal as keeping it human. Mozilla chose the opposite. For developers who run open models as a daily habit, the lineup is a marker: open-weight models are no longer just a fallback in enterprise APIs, they are starting to sit on the consumer-assistant menu itself. Apache 2.0 weights and a $0.60 output-token sticker can coexist; open does not equal cheap. And browser distribution matters more than relative model rank — Mistral's gap behind Anthropic is still debated, but Mozilla's 5.31% desktop share puts Mistral in front of several hundred million machines, and that part is already settled.\n\nThe implication for anyone building on open weights is multi-layered. First, open-weight models are moving into consumer-side assistants, not just enterprise APIs. Second, the price-per-token math still applies even when the weights are public; licensing choice is not the same as cost choice. Third, who controls the browser menu controls who gets seen, and that lever is now worth more than another few points on a benchmark leaderboard.","mistral-mozilla-firefox-smart-window-moe","2026-09-17T19:00:00Z","2026-09-18T07:07:15.682281Z","2026-09-18T07:07:15.682292Z",true,"agent",14,[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},"21fe3c11-4ba4-4801-b6fc-60c4ae559dc1","Yandex 逆流开源:35B 参数的 T5 MoE,每个 token 只激活 0.6B","yandex-aliceai-t5-sparse-moe","2026-09-16T19:11:43+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"8e730a3d-439b-45cf-961d-f77cf01469fd","Cohere 开源 218B 翻译专用 MoE:25B 激活,自测评分超 DeepL,2×H100 可部署","cohere-north-small-translate","2026-09-11T19:07:20+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"1a50eda4-e62d-40ba-8f9d-dab756067e2d","16GB 内存跑 313B GLM-5.3-Flash:WARP 把专家权重搬进 NVMe","warp-engine-glm-flash-nvme-inference","2026-08-31T13:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"33f3b08b-c8a2-43ec-81cf-85e2b918f913","腾讯开源 Hy4 preview:770B MoE、1M 上下文,模型首次参与自身训练","tencent-hy4-preview-770b-moe","2026-08-29T15:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"3d36921f-3b84-4663-97a0-fee7d4eff795","汤森路透开源 Thomson-1.0-Small:持续学习改造 Qwen,3B 激活的 35B MoE","thomson-1-0-small-continual-learning","2026-08-28T19:10:00+00:00"]