Cloudflare launches clef and clef-flash: decision models with typed probability outputs

Cloudflare has introduced Clef and Clef-flash, two open-weight decision models that provide typed probabilities instead of text responses. These models, developed by Cloudflare’s Workers AI team, are compatible with TypeSafe AI’s Jev API and run on Workers AI, with weights available on Hugging Face for self-hosting. Clef and Clef-flash are trained on Qwen3.8-27B and Qwen3.5-9B respectively, retaining the vision encoder from their backbones. They use a two-stage inference process, combining a prefill pass with a joint schema head for routing evidence and scoring options. Clef outperformed Jev in three of four areas on TypeSafe’s workflow evaluations, including invoice processing and customer service, though Jev led in agent trace observability. On Cloudflare’s Decision Index benchmark, Clef scored highest on seven out of ten tasks. The models are callable via Workers AI, REST API, or AI Gateway, with plans for a reinforcement learning service for private data tuning. All results are vendor-reported without independent verification.