Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text
https://www.marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/📌 Cloudflare 首款自研模型不是聊天機器人,是會回機率的決策模型
TL;DR：Cloudflare Workers AI 推出 Clef / Clef-flash,回傳的是機率而不是文字,相容 TypeSafe AI 的 Jev API。
一家以邊緣運算與 CDN 聞名的公司,第一次自己訓練模型,結果訓出來的東西連一句完整的話都不說——這正是 Cloudflare 這次釋出的 Clef 與 Clef-flash 的設計目標。
🤔 跟上 System One 浪潮,但走自己的訓練路線
TypeSafe AI 在 2026 年 9 月 15 日發表了首款「System One」模型 Jev,隨後 Kev-9B、Laya 等開放模型跟進。Cloudflare 這次推出的 Clef 系列,同樣採用 System One API,從 Jev 切換過去只需要換 endpoint 與模型名稱。Clef 讀取輸入的 state 與一組 typed questions,針對每個允許的答案都回傳一個機率,完全沒有自由格式文字輸出。它支援 3 種問題類型,單次請求最多可帶 64 個問題與 4 張圖片。兩個模型都以 Apache 2.0 開放權重,目前都已經可以在 Workers AI 上直接呼叫,權重也同步上架 Hugging Face 供自架部署。
🧩 雙 backbone、兩階段推理、joint schema head
Clef 是從 Qwen3.8-27B 後訓練而來,Clef-flash 則來自 Qwen3.5-9B,兩者都保留了原本 backbone 的視覺編碼器。推理分兩階段:backbone 先對 state 與問題做一次純 prefill(prefill-only)的前向傳播;接著一個小型 transformer——joint schema head——讀取最終的隱藏狀態,把證據路由到每個問題上,讓各欄位之間可以 cross-attend,並對所有選項做聯合評分,最後用每題各自的 softmax 把 logits 轉成機率。訓練時兩個 backbone 都被凍結,只用 rank-256 的低秩適配器(LoRA)聯合最佳化這個 routing head。損失函數結合了 label-smoothed cross-entropy 與用於校準的 Brier loss,另外還有一個次要目標——Reinforcement Learning for Calibrated Decisions(RLCD),會對相鄰的 ordinal 選項給予部分分數。
📊 十項基準中贏七項,但 Jev 在推理類題目仍佔優
在 Cloudflare 從 Decision Index 0.2.1 套件中挑出的 10 項基準裡,Clef 系列在其中 7 項拿到最高分。不過 Jev 在部分推理類題目上仍明顯領先:GPQA Diamond 78.3 分 vs 48.0 分,MMLU-Pro 82.7 vs 65.9,BBH 92.9 vs 73.7。在 TypeSafe 自家的工作流程評測中,Clef 在 4 個領域裡贏了 3 個,差距不大:invoice processing 64.7 vs 61.8,customer service 76.3 vs 76.0,security incidents 62.9 vs 61.7;Jev 則在 agent trace observability 上領先,71.6 vs 68.5。在 Cloudflare 的威脅情資分類情境裡,Clef 完成一次網域分類只花 2.2 秒,相較 gpt-oss-120b 的 4.7 秒快了不少。需要提醒的是,以上數字全部是廠商自行回報,目前還沒有第三方的獨立驗證。
💡 呼叫方式與自架選項都已就位
兩個模型都能透過 Workers AI binding(env.AI.run())、REST API,或 AI Gateway 呼叫。想自架的話,模型卡上標註的測試環境是單張 H200、BF16 權重。Cloudflare 同時宣布了一項強化學習服務,用來在私有資料上微調 Clef,初期先對 forward-deployed engineers 開放,之後才會推出自助式平臺;整條 pipeline 串接了 AI Gateway、Workers AI、Containers,以及一個新的 Trainer 元件,有興趣的團隊可以透過 design partner 表單申請。
⚠️ 數字都是廠商自報,複雜推理仍不是它的主場
Clef 在多項工作流程評測上贏過 Jev,但差距多半是個位數百分點,且評測數字全部來自 Cloudflare 與 TypeSafe 自家測試,尚未有獨立複測。在需要高階推理的題型(GPQA、MMLU-Pro、BBH)上,Jev 仍有明顯優勢,這也呼應了決策模型普遍「不適合複雜推理」的定位。
🎯 實務啟示
如果你已經在用 System One API 串接 Jev 類模型做分流、分類或工具參數檢查,Clef 系列幾乎是零成本的替換選項——介面相容、權重開放、還能接上 Workers AI 的基礎設施;但涉及高階推理判斷的場景,仍建議保留原本的大模型路徑。
🔗 來源
- 標題：Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text
- 作者／機構：Asif Razzaq, MarkTechPost
- 連結：https://www.marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/
#Cloudflare #WorkersAI #DecisionModel #SystemOneAI #OpenWeight #Qwen #ModelCalibration #EdgeAI #LLMOps #MachineLearning
原始資料 MarkTechPost · 收集於 2026-10-02
摘要原文
Cloudflare has released Clef and Clef-flash , the first models trained by its Workers AI team. They are decision models, not chatbots. Each reads an input state and a schema of typed questions. It returns a probability for every allowed answer, with no free-form text. Both are open-weight under Apache 2.0 and compatible with TypeSafe AI’s Jev API. Is it deployable? Yes, Both models run today on Workers AI , and the weights are on Hugging Face for self-hosting. An LLM generates tokens one at a time, and its output still needs parsing. A decision model only answers a fixed set of questions about an input. Clef supports 3 question types: On Workers AI, 1 request carries up to 64 questions and up to 4 images. TypeSafe AI launched Jev , its first ‘System One’ model, on September 15, 2026. Open alternatives like Kev-9B and Laya followed. Clef uses the same System One API. Switching from Jev means changing the endpoint and model name. Clef is post-trained from Qwen3.8-27B , and Clef-flash from Qwen3.5-9B . Both keep the backbone’s vision encoder. Inference has 2 stages. The backbone first runs a single prefill-only pass over the state and questions. A small transformer, the joint schema head, then reads the final hidden states. It routes evidence to each question, lets fields cross-attend, and scores all options jointly. A per-question softmax turns logits into probabilities. Training froze both backbones and jointly optimized the routing head with rank-256 low-rank adapters. The loss pairs label-smoothed cross-entropy with a Brier loss for calibration. A secondary objective, Reinforcement Learning for Calibrated Decisions (RLCD), gives partial credit to adjacent ordinal choices. On Cloudflare’s 10-benchmark shortlist from the Decision Index 0.2.1 suite, a Clef model scored highest on 7. Jev keeps clear leads elsewhere. The full model card shows Jev ahead on GPQA Diamond (78.3 vs 48.0). It also leads MMLU-Pro (82.7 vs 65.9) and BBH (92.9 vs 73.7). On TypeSafe’s own workflow evals , Clef beat Jev in 3 of 4 areas, by small margins. Invoice processing was 64.7 vs 61.8, customer service 76.3 vs 76.0, and security incidents 62.9 vs 61.7. Jev leads agent trace observability, 71.6 vs 68.5. In Cloudflare’s threat intelligence workflow, Clef classified a domain in 2.2 seconds. gpt-oss-120b took 4.7 seconds. All numbers are vendor-reported, with no independent replication yet. *Cloudflare’s internal Decision Index run. All 5 implement the System One API. Sources: Clef docs , Clef-flash docs , Clef card , Jev post , Kev-9B card , Laya card . Both models are callable through the Workers AI binding ( env.AI.run() ), the REST API, or AI Gateway . For self-hosting, the model cards list testing on a single H200 with BF16 weights. Cloudflare also announced a reinforcement learning service for tuning Clef on private data. It starts with Cloudflare’s forward-deployed engineers, with a self-serve platform later. The pipeline combines AI Gateway, Workers AI, Containers and a new Trainer component. Teams can apply via the design partner form . Check out the Model weight , Demo and Technical details . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成