[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-clef-omni-multimodal-decision-model":3,"topics-all":38,"news-related-63ae3eda-4994-4f9f-a812-489d9831a9ca":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},"63ae3eda-4994-4f9f-a812-489d9831a9ca","Clef-omni 开源:决策模型学会听音看片","Cloudflare 开源 Clef-omni 决策模型:基于 Qwen3-Omni-30B-A3B,单次调用同时给音频、视频、图像、文本打校准置信分,21 秒带声视频约 1.5 秒判完。Clef-flash 同步降价至 0.038 美元每百万 token,低于 Jev;文本决策中位约 130 毫秒。","Cloudflare 上周刚开放 Clef 决策模型的权重,10 月 9 日就端出第二弹:Clef-omni 把音频(wav\u002Fmp3)和视频(mp4\u002Fwebm)与文本、图像纳入同一次 API 调用,直接输出结构化决策分数。决策模型这个品类自 TypeSafe 的 Jev 问世以来基本停留在纯文本,Cloudflare 把它推向了「多感官」方向。\n\n## 一次 prefill,四种模态\n\nClef-omni 的底座是 Qwen3-Omni-30B-A3B-Instruct MoE。Cloudflare 保留了它的多模态理解主干,丢弃了文本转语音输出组件——决策模型不是 LLM,不生成输出 token,只在 prefill 阶段对完整载荷打一遍分:每个候选项先从输入里收集关键证据(无论埋在文本、画面还是音轨),再经两阶段注意力路由对全上下文交叉注意,输出置信度。官方示例是设备验收:一张设备照片、一段运转录音、一段风扇视频,配三个问题——铭牌可见吗、声音正常吗、风扇在转吗——一次调用全部判完。训练配方沿用 Clef:冻结 Qwen3 主干、训 LoRA 适配器、标签平滑交叉熵加 Brier 分数校准。权重开放在 HuggingFace。\n\n## 百毫秒级打分,和一轮降价\n\n速度数字相当激进:纯文本决策中位约 130 毫秒,图像约 150 毫秒,音频几百毫秒,21 秒带声视频约 1.5 秒判完。定价方面,Clef-omni 首发 0.15 美元每百万输入 token;同场 Clef-flash 从 0.09 美元砍到 0.038 美元,官方称已低于 Jev。砍价的代价是上下文:Clef-flash 托管版窗口从 64k 收到 24k——官方数据称只有 0.24% 的请求超过 24k,自托管权重仍支持 256k。Clef 本体迁移到 SGLang 服务,官方口径提速约 1.7-2 倍,相关 PR 将随 SGLang 0.5.22 发布。\n\n## 跑分赢在分类,复杂工作流仍让位\n\n官方基准表里,Clef-omni 在 BANKING77(94.8 macro-F1)、CLINC150+OOS(97.7)、Amazon ESCI(57.8)三项分类任务上位列家族第一;但 Home appliances 69.3 大幅落后 Clef-flash 的 97.73,When2Call 63.3 落后 Jev 的 80.97;TypeSafe 自家 evals 里发票处理精确动作 60.2 也低于 Clef 的 64.7。多模态的广度换掉了部分任务深度,这笔交换值不值,取决于工作流里音视频占多大比重。\n\n## 所以呢\n\n两件事值得留意。其一,决策模型正沿着 LLM 的路径多模态化,而 Cloudflare 从「周五晚决定做、周末训练、周四上线」到本周的多模态加降价提速三连,示范了边缘网络厂商用开放权重加超低价切推理市场的节奏。其二,Qwen 开放权重基座又添一个海外商用案例——同一周内,微软的 Microsoft-Decision-1 也构建在 Qwen3.5-9B 之上(Solidot 报道)。当一次「判断」便宜到 0.038 美元每百万 token,值得问的可能不再是「要不要接 AI」,而是「流程里还有多少判断在由人做」。\n\n参考:Cloudflare 官方公告 https:\u002F\u002Fblog.cloudflare.com\u002Fclef-faster-cheaper-multimodal\u002F","https:\u002F\u002Fblog.cloudflare.com\u002Fclef-faster-cheaper-multimodal\u002F","016517e1-7a07-4d85-a702-cbf30d014b76",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"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},"be73b3f7-db47-44d5-adb1-babec1859bbd","en","Clef-omni: Cloudflare's decision model now hears and watches","Cloudflare's open Clef-omni scores audio, video, image and text on Qwen3-Omni-30B-A3B, judging a 21s clip in 1.5s. Clef-flash drops to 0.038 USD\u002FM tokens.","Cloudflare shipped the second act of its open-weight decision model family on October 9: Clef-omni takes audio (wav\u002Fmp3) and video (mp4\u002Fwebm) alongside text and image in a single API call, returning structured decision scores. The decision-model category has been largely text-only since TypeSafe's Jev debuted; Cloudflare is now pushing it toward multi-sensory input.\n\n## One prefill pass, four modalities\n\nClef-omni is built on the Qwen3-Omni-30B-A3B-Instruct MoE foundation. Cloudflare kept its multimodal comprehension backbone and discarded the text-to-speech output components — decision models are not LLMs, generate no output tokens, and run a quick prefill pass over the full payload: every candidate option gathers key evidence from the input (whether buried in text, visuals, or audio streams), then field vectors cross-attend across the context via two-stage attention routing to compute confidence scores. The official example is equipment acceptance: one photo of the unit, an audio recording of it running, and a video of the fan, paired with three questions — is the label visible, does it sound normal, is the fan running — all scored in one call. Training follows the Clef recipe: freeze the Qwen3 backbone, train LoRA adapters, label-smoothed cross-entropy plus Brier score calibration. Weights are open on HuggingFace.\n\n## Sub-150ms scoring and a price cut\n\nThe speed numbers are aggressive: text-only decisions return in about 130 ms median, images in about 150 ms, audio in a few hundred milliseconds, and a full 21-second video clip with sound is scored in about 1.5 seconds — all in one API call. On pricing, Clef-omni launched at 0.15 USD per million input tokens; in the same post Clef-flash dropped from 0.09 to 0.038 USD, which Cloudflare says is now cheaper than Jev. The trade-off for the cut is context: the hosted Clef-flash window shrank from 64k to 24k tokens — official data says only 0.24% of requests exceed 24k, and the self-host weights still support 256k. Clef itself moved to SGLang serving, with official figures of roughly 1.7-2x speedups, and the upstream PR lands in SGLang 0.5.22.\n\n## Wins on classification, yields on complex workflows\n\nIn the official benchmark table, Clef-omni ranks first in the family on BANKING77 (94.8 macro-F1), CLINC150+OOS (97.7), and Amazon ESCI (57.8) classification tasks. But Home appliances scores 69.3, far behind Clef-flash's 97.73; When2Call sits at 63.3 versus Jev's 80.97; and on TypeSafe's own evals, invoice-processing exact actions reach 60.2 versus Clef's 64.7. Multimodal breadth trades away some task depth — whether that trade is worth it depends on how much audio and video your workflow actually carries.\n\n## So what\n\nTwo things stand out. First, decision models are going multimodal along the same path LLMs did, and Cloudflare's cadence — \"decided Friday evening, trained over the weekend, launched Thursday\" to this week's multimodal-plus-price-cut-plus-speedup triple — shows how an edge network vendor uses open weights and aggressive pricing to cut into the inference market. Second, the Qwen open-weight base gained another overseas commercial case — in the same week, Microsoft's Microsoft-Decision-1 was built on Qwen3.5-9B (per Solidot). When a \"judgment\" costs 0.038 USD per million tokens, the question is no longer \"should we plug in AI\" but \"how many judgments in our process are still made by humans.\"\n\nReference: Cloudflare's official announcement https:\u002F\u002Fblog.cloudflare.com\u002Fclef-faster-cheaper-multimodal\u002F","clef-omni-multimodal-decision-model","2026-10-12T15:20:00Z","2026-10-12T15:10:24.430795Z","2026-10-12T15:10:24.430810Z",true,"agent",96,[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},"bbe8d55a-1069-42fa-a342-d944d53fdb4b","操作电脑的 27B 开源权重模型 Holo4:最强版禁商用","holo4-open-weight-computer-use","2026-10-02T13:12:01+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"dc61debd-1d77-4d5a-9d59-5b23c3da07de","蚂蚁开源Realtime-Venus：9B全双工模型边说边干活，三项续聊指标超GPT-4o","ant-realtime-venus-full-duplex-delegation","2026-09-30T23:10:53+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"08872b65-a25f-41f2-877c-a6f265e3c72f","LFM2.5-24B开源:2B激活跑赢Qwen3.6-35B","liquid-ai-lfm2-5-24b-a2b","2026-10-14T23:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"5d4acf64-67d8-404b-9d3f-cdf7761fb8ac","小鹏开源 GroundingPI:4B 视觉 grounding 模型自评超 GPT-6 Astra","groundingpi-4b-visual-grounding-foundation-model","2026-10-13T17:10:26+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"c53a7c3b-7a0c-4cc6-822e-aa0d66e01abe","EmbeddingGemma 2 开源:740M 多模态嵌入跑进手机","embeddinggemma-2-multimodal-embeddings","2026-10-11T21:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"df4159eb-60c1-4267-835b-0932d2a1ac72","d1-3B 开源：零token决策，8毫秒应答","liquid-ai-d1-open-edge-decision-models","2026-10-11T19:14:50+00:00"]