[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-cloudflare-clairic-extraction-model":3,"topics-all":38,"news-related-4c74f8b3-0a8d-4e00-9a7f-10ae44a3c004":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},"4c74f8b3-0a8d-4e00-9a7f-10ae44a3c004","Clairic抽取模型:不生成token,一遍前向出答案","Cloudflare 在 Workers AI 预览上线抽取模型 Clairic:单次前向传播读输入,不生成输出 token,直接返回带置信度的强类型值,支持从文本和图片抽取、计算、估计和预测数值,每百万输入 token 0.1 美元。官方级联实测:比全 LLM 管线快 3.2 倍、省 49%。开放权重计划稍后放出。","把一封投诉邮件变成结构化数据,今天最常见的做法是丢给 LLM 让它吐 JSON。Cloudflare 给出了另一种答案:10 月 19 日,它在 Workers AI 上线预览版模型 Clairic,这款模型从头到尾不生成一个 token,一次前向传播直接给出答案——而且答案是一个带类型的值,不是一段需要再解析的文字。\n\n## System One 谱系的第三块拼图\n\nClairic 的概念源头是 TypeSafe 一个月前提出的 System One 模型:借卡尼曼《思考,快与慢》的框架,让模型做「快而准」的那部分,把「慢推理」留给 LLM。Cloudflare 先前发布的决策模型 Clef 负责在给定选项里做选择,而 Clairic 开辟了新类别——抽取模型,官方称它是第一款 System One 抽取模型。它的定位很具体:发票总额是多少、哪天送达、维修费估多少,这些每天海量涌向大模型的结构化抽取调用,正是它要接管的活。\n\n## 一次前向,字段全并行\n\n调用方式是声明式的:传入 state(文本,最多附 4 张图),再声明一组带类型的 fields——text、number、date、duration、list、estimate、forecast。模型对每个字段返回一个该类型的值:文本给出字符偏移 span,数字遵循类型规则(金额两位小数),estimate 给出中位数和 80% 区间,forecast 逐期给出预测区间。每个答案附带 p_null(值不在输入中的概率)和置信度;找不到就返回 null,而不是编一个像样的猜测。上下文窗口 65,536 token,字段数量无上限,单遍放不下的字段按 8,192 token 上限拆成并行 passes 再合并。官方邮件示例里,649.00 美元货款加 52.10 美元税被同遍算出 701.10 美元,交付日到写信日的 11 天间隔也是从两个日期推出来的;邮件里不存在的追踪号则老实返回 null。\n\n## 级联实测:3.2 倍速、省 49%\n\n官方用一条商品破损理赔管线做对照:Clairic 抽值、Clef 决策、GLM-5.3 写回复,对阵 GLM-5.3 全包。级联版 19.3 秒完成,全 LLM 版 61.7 秒;每百万条理赔成本从 11,740 美元降到 6,032 美元。两个细节更有意思:GLM-5.3 每次跑出的维修费估算都在漂移(250、250、220 美元),Clairic 的估算是稳定的;而全 LLM 管线里 Gemma 4 和 Qwen 直接失败了——推不出维修费这类值。\n\n预测能力也有真实数据背书:官方用 TSA 安检人数做了 21 天滚动预测,平均误差 3.3%,21 天里 17 天的实际值落在 80% 区间内,还从纯数字里学出了「周二周六淡、周四五日旺」的周期。\n\n## 边界清晰,姿态克制\n\n官方没有回避局限:多步算术还不行,长链条计算交给代码;估计需要输入里有证据;预测只是基础版,不能替代专用时序模型;校准是测出来的,不保证。定价每百万输入 token 0.10 美元,只收输入——因为它不吐输出 token。开放权重版本承诺「soon」。\n\n我的看法:Clairic 代表的趋势比模型本身更重要——当一类调用形状足够稳定(字段固定、类型固定、量大),就值得从通用 LLM 里剥离出来做成专用单遍模型,把成本和延迟打到地板上。Jev、Clef、Clairic 接连出现,说明「非 LLM 的小模型」正在从概念变成产品线。对开发者,值得问的问题是:你今天发给大模型的调用里,有多少其实只是要一个数字?\n\n原文:https:\u002F\u002Fblog.cloudflare.com\u002Fclairic-extraction-model\u002F\n","https:\u002F\u002Fblog.cloudflare.com\u002Fclairic-extraction-model\u002F","e73c704c-6268-4e8d-b1d1-7873aea8124d",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",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},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"6383abe3-0447-453a-bbdf-100ced9023a3","en","Cloudflare Clairic: one forward pass, zero output tokens","Cloudflare previews Clairic, a one-pass extraction model on Workers AI: typed values with confidence, no output tokens, $0.10 per million input tokens.","Turning a complaint email into structured data usually means asking an LLM to spit out JSON. Cloudflare has a different answer: on October 19 it shipped Clairic as a preview on Workers AI, a model that generates no tokens at all — one forward pass, and the answer comes out as a typed value instead of text you have to parse again.\n\n## The third piece of the System One lineup\n\nThe concept traces back to System One models, which TypeSafe introduced a month ago, borrowing Daniel Kahneman's fast-and-slow framing: let a model handle the fast, precise part and leave deliberate reasoning to LLMs. Cloudflare's earlier decision model Clef picks among given options; Clairic opens a new class — extraction models — and Cloudflare calls it the first of that class. Its target is concrete: the flood of structured extraction calls that today go to frontier LLMs — invoice totals, delivery dates, repair cost estimates.\n\n## One forward pass, all fields in parallel\n\nThe API is declarative: pass a state (text, plus up to four images) and a map of typed fields — text, number, date, duration, list, estimate, forecast. Each field returns a value of its type: text answers carry character-offset spans, numbers follow type rules (two decimals for cents), estimates return a median and an 80% interval, forecasts return one interval per step ahead. Every answer ships with p_null (the probability the value isn't in the input) and a confidence score; when the value isn't there, the model returns null rather than a plausible guess. The context window is 65,536 tokens, there is no limit on field count, and fields that don't fit beside the state in one 8,192-token pass are split into parallel passes and merged. In the official email example, a $649.00 price plus $52.10 tax was summed to $701.10 in the same pass, the 11 days between delivery and the email was worked out from two dates, and a tracking number that never appeared came back null.\n\n## Cascade test: 3.2x faster, 49% cheaper\n\nCloudflare benchmarked a damaged-product claim pipeline: Clairic extracts, Clef decides, GLM-5.3 writes the reply, versus GLM-5.3 doing everything. The cascade finished in 19.3 seconds against 61.7; cost per million claims fell from $11,740 to $6,032. Two details stand out: GLM-5.3's repair-cost estimate drifted on every run ($250, $250, $220) while Clairic's stayed stable, and in the all-LLM pipeline both Gemma 4 and Qwen failed outright — they couldn't deduce the repair cost.\n\nForecasting has real-data backing too: over 21 days of rolling forecasts on TSA checkpoint counts, average error was 3.3%, actuals landed inside the 80% interval on 17 of 21 days, and the model picked up the weekly rhythm — quiet Tuesdays and Saturdays, busy Thursdays, Fridays and Sundays — from raw numbers alone.\n\n## Clear limits, restrained posture\n\nCloudflare is candid about weaknesses: multi-step arithmetic isn't there yet (let code do the math), estimates need evidence in the input, forecasting is basic and not a replacement for dedicated time-series models, and calibration is measured, not guaranteed. Pricing is $0.10 per million input tokens, with no output-token charge — because there are no output tokens. Open weights are promised \"soon\".\n\nMy take: the trend Clairic represents matters more than the model itself. When a class of calls is stable enough — fixed fields, fixed types, high volume — it pays to peel it off the general LLM into a purpose-built one-pass model and drive cost and latency to the floor. Jev, Clef and Clairic appearing in quick succession suggests that \"small non-LLM models\" are turning from concept into a product line. The question worth asking: of the calls you send to an LLM today, how many only ever wanted a number?\n\nSource: https:\u002F\u002Fblog.cloudflare.com\u002Fclairic-extraction-model\u002F\n","cloudflare-clairic-extraction-model","2026-10-22T23:10:00Z","2026-10-22T23:07:58.981524Z","2026-10-22T23:07:58.981538Z",true,"agent",30,[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},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","minimax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b667e52f-ec7d-4ca4-8d9e-1db81e1a5616","DeepSeek论文:890字节KV缓存的三层架构账","deepseek-v41-flash-kv-cache-paper","2026-09-18T15:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"823e17ef-5927-40e6-9efd-08c4958922f0","DeepSeek 旗舰 V4-Pro 今日退役:552B 的 V4.1-Flash 全面接班","deepseek-v4-pro-retires-flash-routes","2026-09-14T15:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"550cee5e-18e8-4236-9304-7207ebc221a8","Agnes 3.0 Flash 开源:72 层仅 18 层带 KV 缓存,33B 单卡跑 262k 上下文","agnes-3-0-flash-preview-open-weights","2026-09-13T15:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"1bf979b5-2b52-4a19-8e2d-e0a549bb24ba","fal 后训练版 MiniMax H3:5 秒视频约 3 秒生成,吞吐 35 倍","fal-h3-max-post-trained-video","2026-08-29T23:02:12+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00"]