Cloudflare open-sources Clef decision models
Cloudflare has released Clef and Clef-flash, two open-weight decision models built to give agents fast, structured, probability-scored classifications, and is pitching them as faster than Typesafe AI's Jev.
Cloudflare announced two decision models, Clef and Clef-flash. Both are hosted on Workers AI and published on Hugging Face under an Apache 2.0 license. The company also introduced a reinforcement learning product that lets customers fine-tune Clef. The post does not give pricing or details of how that product works.
Decision models return typed answers with probabilities rather than open-ended text, so code can act on the result directly, for example to route a support ticket or escalate it to a human. Cloudflare says Clef is compatible with the API of Typesafe AI's Jev System One. Clef adds a vision encoder, which Jev lacks because it handles text only, and a 64k context window, double Jev's 32k.
Cloudflare's own benchmarks show mixed but mostly favourable results. Clef beats Jev on BANKING77 (94.20 vs 79.74 macro-F1) and CLINC150+OOS (97.43 vs 89.27). Jev leads on When2Call, BRIGHT and agent-trace observability. A DiffusionGemma variant of Jev tops PhishNChips. Cloudflare says Clef leads the Jev Decision Index.
The company puts the clearest gap in latency. Across 43 evals, Clef-flash posted a median of 38.8 ms against Jev's 524.1 ms. Only Laya was faster, at 5.8 ms, but it scored far lower on quality.
Internally, Cloudflare's Threat Intelligence team uses Clef to categorise domains. Fetching, rendering and classifying a site took 2.2 seconds, compared with 4.7 seconds for gpt-oss-120b in the same workflow.
Nobody outside Cloudflare has verified these figures yet. The available text of the post ends before it describes the RL platform in detail.