<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e" -->

---
title: Introducing Clef: our open-source decision models, and...
description: Cloudflare has released Clef and Clef-flash, open-source decision models hosted on Workers AI, designed for fast, deterministic structured classification...
canonical: https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Introducing Clef: our open-source decision models, and new RL fine-tuning platform | daily.dev
og:description: Cloudflare has released Clef and Clef-flash, open-source decision models hosted on Workers AI, designed for fast, deterministic structured classification...
og:url: https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e
og:image: https://api.daily.dev/og/posts/oTMxhXg4E.png
og:image:alt: Introducing Clef: our open-source decision models, and new RL fine-tuning platform
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.

# Introducing Clef: our open-source decision models, and new RL fine-tuning platform

**[Cloudflare](https://daily.dev/sources/cloudflare)** · 12 min read · 0 upvotes · 0 comments

## Summary

Cloudflare has released Clef and Clef-flash, open-source decision models hosted on Workers AI, designed for fast, deterministic structured classification useful in agentic workflows. Built on a Qwen backbone with a non-autoregressive, prefill-only scoring approach, Clef claims higher accuracy and lower latency than Typesafe AI's Jev model and general-purpose LLMs like gpt-oss-120b across dozens of benchmarks. The models are Jev-API compatible, licensed under Apache 2.0, and downloadable from Hugging Face. Cloudflare is also launching a reinforcement learning fine-tuning service, starting with a hands-on forward-deployed engineer team and eventually a self-serve platform built on AI Gateway, Containers, and a new Trainer component for redeploying fine-tuned models on Workers AI.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://blog.cloudflare.com/clef-decision-models>

## Questions this post answers

### What is Cloudflare's Clef decision model and how does it compare to Typesafe AI's Jev?

Clef and Clef-flash are Cloudflare-trained decision models hosted on Workers AI that produce typed, probability-based classifications and are fully API-compatible with Jev. Clef has a vision encoder for image classification (Jev is text-only), a 64k context window versus Jev's 32k, and beat Jev in 3 of 4 evals on Typesafe's own eval suite, with lower median latency across 43 benchmarks.

_Teams weighing decision-model options for agentic pipelines can track comparisons like this on daily.dev._

### How much faster is Cloudflare's Clef model than a general-purpose LLM for domain classification?

In Cloudflare's Threat Intelligence domain classification workflow, Clef took 2.2 seconds to fetch, render, and classify a website, versus 4.7 seconds for gpt-oss-120b, which also returned only two classifications compared to Clef's broader output. This roughly doubles throughput for agentic pipelines needing fast, structured decisions.

_Developers optimizing agent latency can follow benchmarks like this on daily.dev._

### What base model and training approach does Cloudflare use to build the Clef decision model?

Clef is built on a Qwen backbone (Qwen3.8-27B for Clef, Qwen3.5-9B for Clef-flash) that is frozen while jointly optimizing a routing head with rank-256 low-rank adapters. Training uses label-smoothed cross-entropy plus Brier loss for probability calibration, synthetic datasets with permuted field orders, and a custom reinforcement learning method called RLCD for calibrated, non-autoregressive decision scoring.

_Engineers evaluating custom classifier training approaches can dig into details like this on daily.dev._

## Similar posts on daily.dev

- [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
- [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

---

Tags: [#open-source](https://daily.dev/tags/open-source), [#cloudflare](https://daily.dev/tags/cloudflare), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning)

[View this post on daily.dev](https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e)

```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":"Introducing Clef: our open-source decision models, and new RL fine-tuning platform","url":"https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e"},"datePublished":"2026-10-01T15:35:17.940Z","dateModified":"2026-10-02T01:33:33.675Z","description":"Cloudflare has released Clef and Clef-flash, open-source decision models hosted on Workers AI, designed for fast, deterministic structured classification...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/7ed68c578f7b64d571006d09b66f84d2?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/7ed68c578f7b64d571006d09b66f84d2?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Cloudflare","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":"Cloudflare","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/38522e1d11354cd6b7af66f9d4316735","url":"https://daily.dev/sources/cloudflare"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,cloudflare,reinforcement-learning","timeRequired":"PT12M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Cloudflare","item":"https://daily.dev/sources/cloudflare"},{"@type":"ListItem","position":3,"name":"Introducing Clef: our open-source decision models, and new RL fine-tuning platform"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/introducing-clef-our-open-source-decision-models-and-new-rl-fine-tuning-platform-otmxhxg4e#faq","mainEntity":[{"@type":"Question","name":"What is Cloudflare's Clef decision model and how does it compare to Typesafe AI's Jev?","acceptedAnswer":{"@type":"Answer","text":"Clef and Clef-flash are Cloudflare-trained decision models hosted on Workers AI that produce typed, probability-based classifications and are fully API-compatible with Jev. Clef has a vision encoder for image classification (Jev is text-only), a 64k context window versus Jev's 32k, and beat Jev in 3 of 4 evals on Typesafe's own eval suite, with lower median latency across 43 benchmarks. Teams weighing decision-model options for agentic pipelines can track comparisons like this on daily.dev."}},{"@type":"Question","name":"How much faster is Cloudflare's Clef model than a general-purpose LLM for domain classification?","acceptedAnswer":{"@type":"Answer","text":"In Cloudflare's Threat Intelligence domain classification workflow, Clef took 2.2 seconds to fetch, render, and classify a website, versus 4.7 seconds for gpt-oss-120b, which also returned only two classifications compared to Clef's broader output. This roughly doubles throughput for agentic pipelines needing fast, structured decisions. Developers optimizing agent latency can follow benchmarks like this on daily.dev."}},{"@type":"Question","name":"What base model and training approach does Cloudflare use to build the Clef decision model?","acceptedAnswer":{"@type":"Answer","text":"Clef is built on a Qwen backbone (Qwen3.8-27B for Clef, Qwen3.5-9B for Clef-flash) that is frozen while jointly optimizing a routing head with rank-256 low-rank adapters. Training uses label-smoothed cross-entropy plus Brier loss for probability calibration, synthetic datasets with permuted field orders, and a custom reinforcement learning method called RLCD for calibrated, non-autoregressive decision scoring. Engineers evaluating custom classifier training approaches can dig into details like this on daily.dev."}}]}
```

