Cloudflare released Clef and Clef-flash on October 1 for agent workflows that need a bounded decision. A support message can be classified by urgency and destination team, for example, with probabilities attached to the permitted answers.
The models score those choices without writing an intermediate response token by token. Clef is the larger model, while Clef-flash prioritizes lower latency. Cloudflare describes image input and a 64,000-token context window for Clef. Both models use a Jev-compatible API.
The published evaluations show different strengths rather than one model winning every task. Clef-flash is faster in Cloudflare’s latency comparison, while the larger model performs better on several classification and retrieval tests. A workflow still needs to decide what confidence is sufficient to act and when to ask a person.
Both models are hosted on Workers AI, and Cloudflare has released their weights under Apache 2.0. Fine-tuning is initially offered through a hands-on service with its engineering team; the self-service training platform is planned for later.