Cloudflare Releases Open-Source Decision Models Clef and Clef-flash, Debuts RL Fine-Tuning Platform

Cloudflare has released Clef and Clef-flash, two open-source decision models trained by the company and hosted on Workers AI. These models are designed to produce bounded, structured outputs for agentic workflows, offering a deterministic alternative to non-deterministic Large Language Models (LLMs). The models are fully open-sourced on Hugging Face under an Apache 2.0 license and are API-compatible with Jev, allowing for easy integration into existing systems.

Performance and Capabilities

Clef is positioned as a leader on the Jev Decision Index, with Cloudflare reporting that it outperforms Jev in 3 out of 4 areas in Typesafe’s eval suite. Key differentiators include a vision encoder for image classification and a 64k context window, compared to Jev’s 32k. In internal tests, Clef classified website domains in 2.2s, whereas the fastest general LLM, gpt-oss-120b, took 4.7s and returned fewer classifications. The models also demonstrate lower latency than Jev, with Clef-flash achieving a median latency of 38.8ms versus Jev’s 524.1ms.

Architecture and Training

Clef builds upon the concept of exposing logprobs from a large language model to generate deterministic probabilities, but uses Qwen as its base model. Specifically, Clef uses Qwen3.8-27B and Clef-flash uses Qwen3.5-9B. The architecture employs a non-autoregressive decision step where valid schema choices are scored in parallel, avoiding intermediate text generation. Training utilized Reinforcement Learning for Calibrated Decisions (RLCD) to improve accuracy and generalization, alongside label-smoothed cross-entropy and Brier loss.

New RL Fine-Tuning Platform

Cloudflare also debuted a new reinforcement learning (RL) product to help customers fine-tune Clef for specific use cases. The service initially offers hands-on support via a forward-deployed engineer (FDE) team, with plans for a self-serve platform. The workflow leverages existing Cloudflare primitives: AI Gateway for dataset creation, Workers AI for rollouts, Containers for RL sandboxes, and a new Trainer component to update model weights.