Clef: Cloudflare's open-source decision model

LLMs are great at open-ended work, but agents also make lots of small, bounded decisions: is this urgent, which team, how severe. Clef is built only for that, and it's compatible with Typesafe AI's Jev API.

How it works

1. 1What we got. Cloudflare's own example ticket ("Checkout has been failing for every customer for the last hour") on Workers AI: urgent 99.1%, team technical 80.9%, severity Critical 97.0%. Three runs, identical numbers.
2. 2Two sizes. Clef (Qwen3.8-27B base) for precision, Clef-flash (Qwen3.5-9B base) for latency. In Cloudflare's own tests, Clef-flash's median latency was 38.8 ms; Jev's hosted API was 524.1 ms (Cloudflare notes these aren't directly comparable).
3. 3The index. In Cloudflare's own run of the Jev Decision Index, Clef scores 61.21 vs Jev 57.91. These results are self-reported and not on the upstream board; Jev still wins several benchmarks, especially reasoning-heavy ones.
4. 4Images and context. Clef has a vision encoder (up to 4 images per request) and a 64k context window on Workers AI.
5. 5Run it. Workers AI models @cf/cloudflare/clef ($0.24 / M input tokens) and @cf/cloudflare/clef-flash, or download the Apache-2.0 weights from Hugging Face.

Good to know

• Cloudflare also announced a reinforcement-learning fine-tuning service for Clef, starting as a hands-on offering.
• An independent guide by @copilot_shogo. Not affiliated with Cloudflare.