Cloudflare releases open decision models for agent routing
Cloudflare released Clef and Clef-flash, 27-billion- and 9-billion-parameter models that answer typed yes/no, choice and score questions rather than generating free-form text. Both are live on Workers AI, have Apache-2.0 weights and use a 64,000-token context window; a hands-on reinforcement-learning tuning service is open to design partners.
Why it matters: Small structured decision models can move routing, escalation and policy classification out of a general language-model prompt and into a cheaper, inspectable output contract. That makes calibration and abstention quality more important than fluent reasoning.