Cloudflare releases Clef, its first open models, to rival Jev at classification

Cloudflare published Clef and Clef-flash on 1 October, its first models trained by the Workers AI team. According to the company's blog post, Clef has 27 billion parameters and is built on Qwen, while Clef-flash has 9 billion. Both carry a 64,000-token context window and a vision encoder, and the weights are on Hugging Face under Apache 2.0.

They are decision models, not chatbots. Cloudflare says they return bounded, typed outputs with confidence scores, so an agent can classify a request, route it or pick an action without a person in the loop. The Workers AI documentation says they accept text, JSON, images or video and return probabilities across numeric, choice or score questions.

The company claims speed and accuracy. It says Clef leads the Jev Decision Index and reaches a median latency of 209.3 milliseconds against 524.1 for Jev, while Clef-flash manages 38.8. It also reports 94.20 macro-F1 on BANKING77 and 97.43 on CLINC150+OOS. All of these figures come from Cloudflare, and nobody outside the company has repeated them yet.

Pricing is listed at $0.24 per million input tokens for Clef and $0.09 for Clef-flash on Workers AI. Cloudflare says it does not read, store or train on customer requests. It also announced a reinforcement-learning service for fine-tuning decision models on a customer's own data, which starts with forward-deployed engineers and is not yet self-serve.

Clef lands two weeks after TypeSafe AI opened early access to Jev on 15 September, which is closed-weight. Ollama added a Jev-style endpoint to version 0.35 this week, according to its release notes, and several open clones have appeared on GitHub. Cloudflare's own use cases include threat-intelligence domain classification, support-ticket triage and bot detection.