--- title: Cloudflare releases Clef, open-weight decision models,... description: Cloudflare has released Clef and Clef-flash, open-weight decision models hosted on Workers AI, designed for fast, deterministic structured classification used... canonical: https://daily.dev/posts/cloudflare-releases-clef-open-weight-decision-models-and-an-rl-fine-tuning-service-q5kwww262 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Cloudflare releases Clef, open-weight decision models, and an RL fine-tuning service | daily.dev og:description: Cloudflare has released Clef and Clef-flash, open-weight decision models hosted on Workers AI, designed for fast, deterministic structured classification used... og:url: https://daily.dev/posts/cloudflare-releases-clef-open-weight-decision-models-and-an-rl-fine-tuning-service-q5kwww262 og:image: https://api.daily.dev/og/posts/Q5KWww262.png og:image:alt: Cloudflare releases Clef, open-weight decision models, and an RL fine-tuning service og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Cloudflare releases Clef, open-weight decision models, and an RL fine-tuning service **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary Cloudflare has released Clef and Clef-flash, open-weight decision models hosted on Workers AI, designed for fast, deterministic structured classification used in routing and gating decisions within agentic workflows. Built on a Qwen backbone with a non-autoregressive, prefill-only scoring approach, Cloudflare claims the models beat Typesafe AI's Jev model and general-purpose LLMs like gpt-oss-120b on accuracy and latency, though these are self-reported benchmarks. The models are Jev-API compatible, Apache 2.0 licensed, and downloadable from Hugging Face. Cloudflare also launched a reinforcement learning fine-tuning service, starting with forward-deployed engineers before a self-serve platform arrives, built on AI Gateway, Containers, and a new Trainer component. LangChain's Harrison Chase praised the open-weight approach, arguing decision models should be as swappable as the main model in an agent harness. ## Content Cloudflare has released two decision models, Clef and Clef-flash. They are the first models trained by its Workers AI team, and the weights are open under Apache 2.0. You can download them from Hugging Face or run them on Workers AI. ## What they do Decision models handle fast, deterministic structured classification, the kind of small routing and labeling choices that agentic workflows make constantly. Clef is built on a Qwen backbone and uses a non-autoregressive, prefill-only scoring approach, so it scores options instead of generating text token by token. Cloudflare says Clef beats Typesafe AI's Jev model on accuracy and latency across dozens of benchmarks. It makes the same claim against general-purpose LLMs such as gpt-oss-120b. Those are Cloudflare's numbers, and I'd want to see independent runs. The models are Jev-API compatible, which should make switching straightforward for anyone already using Jev. ## Early reaction People are already trying it. @lucataco ran clef-flash locally on an M5 Max, and @kentcdodds passed that along. @steipete wrote that he had "never seen an idea spreading so fast." @hwchase17 called it "decision model season," noted that he had expected more entrants and that this one ships open weights, and credited @ritakozlov and the Cloudflare team. His practical take: a harness should be able to swap its decision model as easily as its main model. I agree with that. ## RL fine-tuning Cloudflare is also launching a reinforcement learning fine-tuning service. It starts with a hands-on team of forward-deployed engineers and is meant to become a self-serve platform. That platform is built on AI Gateway, Containers, and a new Trainer component, which redeploys fine-tuned models on Workers AI. ## Questions this post answers ### What is Cloudflare's Clef model and what is it used for? Clef and Clef-flash are open-weight decision models released by Cloudflare on Workers AI, built on a Qwen backbone using a non-autoregressive, prefill-only scoring approach instead of token-by-token generation. They are designed for fast, deterministic structured classification, such as the routing and gating decisions agentic workflows make constantly. They are Apache 2.0 licensed, Jev-API compatible, and downloadable from Hugging Face. _Developers wiring decision models into agent harnesses can follow releases like Clef on daily.dev._ ### Is Cloudflare's Clef model faster than gpt-oss-120b for classification tasks? Cloudflare claims Clef and Clef-flash deliver higher accuracy and lower latency than general-purpose LLMs such as gpt-oss-120b, as well as Typesafe AI's Jev model, across dozens of benchmarks. These figures come from Cloudflare's own testing rather than independent verification, so they should be treated as vendor-reported claims pending outside validation. _Anyone comparing decision models for routing calls can track independent benchmarks like this on daily.dev._ ### What is Cloudflare's new reinforcement learning fine-tuning service? Cloudflare launched an RL fine-tuning service that initially works through a hands-on team of forward-deployed engineers, with a self-serve platform planned for later. The self-serve version will be built on AI Gateway, Containers, and a new Trainer component that redeploys fine-tuned models directly onto Workers AI. _Teams planning to fine-tune their own agent models can follow services like this as they evolve on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 71 comments across x (as of 2026-10-02). **TL;DR:** Reactions are largely enthusiastic about decision models as a faster, cheaper alternative to prompting LLMs for routing/classification, though several note this is yet another fast clone of an existing idea (Jev) and ask practical questions about latency, use cases, and naming. **Sentiment:** 55% positive · 35% mixed · 10% skeptical **The case for** - Several note decision models answer routing/classification questions far faster and cheaper than generative LLMs, since you don't need prose to pick an enum. - One person reported dropping latency to 20ms and compute cost to near-zero after switching to discrete decision models in production. - Being Jev-API compatible is seen as making it easy to swap in as a drop-in replacement. - The RL fine-tuning platform is flagged by one commenter as potentially the bigger story, since decision models are cheap to train and evaluate. - Open-weight availability is praised as opening up a space previously dominated by closed frontier models. **The pushback** - Multiple people frame this as yet another fast copy of an existing idea (Jev/OpenClaw-style), questioning the originality. - Some find the naming confusing or bad (e.g. how to pronounce 'Clef', comparisons to similarly-named 'Jev'/'Kev'). - A few ask basic clarifying questions about what problem decision models actually solve or how to use them, suggesting unclear communication. - One commenter is skeptical that reliability issues from the earlier Jev-style approach have actually been solved. - Concern that the market is getting flooded with too many similar benchmarked models, with a call for a unified benchmarks portal. **By community** - x (mixed): Replies mix genuine enthusiasm about fast, cheap decision models with a recurring theme that this is yet another quick clone of an existing idea, alongside scattered practical questions and naming jokes. **Hottest debate:** Whether this release represents genuine innovation versus just another fast-follow clone of an existing idea (Jev/OpenClaw). **Open questions** - What is the actual latency per call when running Clef on Workers in production? - What concrete use cases justify adopting a decision model versus existing classifiers or prompted LLMs? - Has the reliability problem noted in the earlier Jev-style approach actually been addressed? **Highlights** > @steipete decision models are spreading fast because generative llms were always the wrong tool for routing: you don't need prose to pick an enum. in our pipeline targeting brazil, discrete decision models dropped latency to 20ms and compute cost to $0. typed control flow wins > — [Kizuno18 on x](https://x.com/Kizuno18/status/2105847759412842745) > @steipete The RL fine-tuning platform may be the bigger story — decision models are cheap to train and evaluate, so RL is tractable here in a way it isn't for frontier LLMs. And a fast open-source 'judge' at the edge is exactly what agent tool-routing needs. > — [ela\_euk on x](https://x.com/ela_euk/status/2105835098755019136) > @steipete So nice to see that the Jev ideia is spreading out, specially in OpenClaw. Sad that you guys still didn't solve the reliability problem. > — [CheriffAI on x](https://x.com/CheriffAI/status/2105807125721096652) > @steipete I still keep chat models off the yes/no gate. Last time I let one “decide” a tool path it wrote a paragraph instead of a pick — and I couldn’t tell what it chose until I re-read the log. > — [bykimdohoon on x](https://x.com/bykimdohoon/status/2105789747860050428) > @urivalev @steipete Clef (27B) and Clef-flash (9B) are open-weight, Jev-API compatible decision models. They lead most benchmarks vs Jev (e.g. BANKING77 F1 94.2 vs 79.7), run 2.5-13x faster (medians 209/39 ms vs 524 ms), support images/video + 64k context (Jev is text-only, 32k state), and cost more > — [grok on x · 1 comments](https://x.com/grok/status/2105838044330807547) **Source threads** - [x](https://x.com/steipete/status/2105778011635400949) · 0 points · 71 comments - [x](https://x.com/kentcdodds/status/2105833403195261008) · 0 points · 0 comments ## Similar posts on daily.dev - [One year later: Sovereign AI and the fight for choice](https://daily.dev/posts/one-year-later-sovereign-ai-and-the-fight-for-choice-xjyth832x) · Cloudflare · 0 upvotes · 0 comments - [Cloudflare’s AI Platform: an inference layer designed for agents](https://daily.dev/posts/cloudflare-s-ai-platform-an-inference-layer-designed-for-agents-70umlv07w) · Cloudflare · 1 upvotes · 0 comments - [Cloudflare Workers AI: run LLMs without API keys](https://daily.dev/posts/cloudflare-workers-ai-run-llms-without-api-keys-sxc7e3lma) · Flavio Copes · 1 upvotes · 0 comments - [Cloudflare Demonstrates Moltworker, Bringing Self-Hosted AI Agents to the Edge](https://daily.dev/posts/cloudflare-demonstrates-moltworker-bringing-self-hosted-ai-agents-to-the-edge-tpslsmm7l) · InfoQ · 2 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#ai-agents](https://daily.dev/tags/ai-agents), [#cloudflare](https://daily.dev/tags/cloudflare), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/cloudflare-releases-clef-open-weight-decision-models-and-an-rl-fine-tuning-service-q5kwww262) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Cloudflare releases Clef, open-weight decision models, and an RL fine-tuning service","url":"https://daily.dev/posts/cloudflare-releases-clef-open-weight-decision-models-and-an-rl-fine-tuning-service-q5kwww262","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/cloudflare-releases-clef-open-weight-decision-models-and-an-rl-fine-tuning-service-q5kwww262"},"datePublished":"2026-10-01T17:38:02.596Z","dateModified":"2026-10-02T03:49:07.348Z","description":"Cloudflare has released Clef and Clef-flash, open-weight decision models hosted on Workers AI, designed for fast, deterministic structured classification used...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/cloudflare-releases-clef-open-weight-decision-models-and-an-rl-fine-tuning-service-q5kwww262","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,ai-agents,cloudflare,reinforcement-learning","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Cloudflare releases Clef, open-weight decision models, and an RL fine-tuning service"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/cloudflare-releases-clef-open-weight-decision-models-and-an-rl-fine-tuning-service-q5kwww262#faq","mainEntity":[{"@type":"Question","name":"What is Cloudflare's Clef model and what is it used for?","acceptedAnswer":{"@type":"Answer","text":"Clef and Clef-flash are open-weight decision models released by Cloudflare on Workers AI, built on a Qwen backbone using a non-autoregressive, prefill-only scoring approach instead of token-by-token generation. They are designed for fast, deterministic structured classification, such as the routing and gating decisions agentic workflows make constantly. They are Apache 2.0 licensed, Jev-API compatible, and downloadable from Hugging Face. Developers wiring decision models into agent harnesses can follow releases like Clef on daily.dev."}},{"@type":"Question","name":"Is Cloudflare's Clef model faster than gpt-oss-120b for classification tasks?","acceptedAnswer":{"@type":"Answer","text":"Cloudflare claims Clef and Clef-flash deliver higher accuracy and lower latency than general-purpose LLMs such as gpt-oss-120b, as well as Typesafe AI's Jev model, across dozens of benchmarks. These figures come from Cloudflare's own testing rather than independent verification, so they should be treated as vendor-reported claims pending outside validation. Anyone comparing decision models for routing calls can track independent benchmarks like this on daily.dev."}},{"@type":"Question","name":"What is Cloudflare's new reinforcement learning fine-tuning service?","acceptedAnswer":{"@type":"Answer","text":"Cloudflare launched an RL fine-tuning service that initially works through a hands-on team of forward-deployed engineers, with a self-serve platform planned for later. The self-serve version will be built on AI Gateway, Containers, and a new Trainer component that redeploys fine-tuned models directly onto Workers AI. Teams planning to fine-tune their own agent models can follow services like this as they evolve on daily.dev."}}]} ```