Announced October 1, 2026, Clef scores schema choices in one forward pass rather than generating text. The full Clef stack uses Qwen3.8-27B frozen plus a rank-256 LoRA and routing head, 64k context, and a vision encoder. Clef-flash uses Qwen3.5-9B for a faster hop. Training uses label-smoothed cross-entropy, Brier loss, and RLCD. Cloudflare's table listed Clef-flash at about 38.8ms median versus Jev at about 524ms; independent Hacker News testers also reported hosted Clef slower than Jev on their workloads. Pricing cited around launch was about $0.24 per million input tokens for Clef and $0.09 for Clef-flash.