Cloudflare launches Clef to speed up programmatic AI decisions

Cloudflare is entering the decision model arena with Clef, a family of fast, open-weight models designed for low-latency classification and routing.

Cloudflare released Clef and Clef-flash under an Apache 2.0 license on Hugging Face, alongside hosted versions on Workers AI. Instead of generating general text like a standard LLM, these models output typed probabilities for quick workflow decisions—like routing support tickets or checking threat data. Cloudflare also introduced a reinforcement learning platform to let developers fine-tune Clef for custom use cases.

Why it matters: Standard LLMs are often too slow and expensive for simple background logic. In Cloudflare's domain classification tests, Clef rendered and labeled sites in 2.2 seconds, compared to 4.7 seconds for a general LLM. With a 64k context window, vision support, and median latencies dropping to 38.8 ms on Clef-flash, it gives autonomous agents a much faster hot path.

Here's the gist: you can run the open weights locally or try them directly on Cloudflare Workers AI.

Expect more infra companies to ship tiny, single-purpose decision models as agent workflows shift toward speed over creative text.