Cloudflare has released Clef and Clef-flash, its first models trained by the Workers AI team. These are decision models that read an input state and a schema of typed questions to return a probability for every allowed answer, rather than generating free-form text.

  • Both models are open-weight under Apache 2.0 and compatible with TypeSafe AI’s Jev API.
  • Clef is post-trained from Qwen3.8-27B, while Clef-flash is based on Qwen3.5-9B; both retain the backbone's vision encoder.
  • The models support three question types: yes/no, choice, and score, with inference using a joint schema head to route evidence and cross-attend fields.
  • On Cloudflare’s Decision Index 0.2.1 suite, Clef scored highest on 7 of 10 benchmarks, outperforming Jev on BANKING77 (94.20 vs 79.74) and CLINC150+OOS (97.43 vs 89.27).
  • Median latency is 209.3 ms for Clef and 38.8 ms for Clef-flash, compared to 524.1 ms for Jev.

The release provides a deployable option for structured decision-making tasks on Workers AI or via self-hosting, with an upcoming reinforcement learning service for fine-tuning on private data.