Cloudflare released Clef and Clef-flash as open-source decision models. They are hosted on Workers AI for high-speed classification and agent workflows. Clef leads the Jev Decision Index among tested models. Both models are Jev-API compatible and open-sourced under Apache 2.0. Cloudflare also launched a reinforcement learning platform for fine-tuning Clef. Clef supports vision and text inputs with a 64k context window. It outputs typed answers with probabilities for programmatic decisions. Clef is faster than Jev and most other decision models in benchmarks. Workers AI hosting adds low-latency edge inference. Clef is enterprise-ready with no data retention for inference. Fine-tuning services are available via Cloudflare’s FDE team.
What changed
Decision models were not widely available as open-source, fast, structured-output classifiers on Cloudflare’s platform.
What it unlocks
Running fast, structured decision workflows in production without retraining for new categories.
- 209.3 ms median latency for Clef
- 38.8 ms median latency for Clef-flash
- 64k context window
- 98.76% BFCL case exact for Clef-flash
What you need to act on it
- a Cloudflare account for Workers AI
- API access for fine-tuning
Sources