Cloudflare released Clef, an open source family of what it calls decision models, running on its Workers AI platform, according to the company's blog. The lineup includes a full-size Clef model and a faster Clef-flash variant.
Decision models classify input and return bounded, structured outputs with confidence scores, rather than generating open-ended text the way a general-purpose large language model does, Cloudflare said. The company's example is reading a customer support message to decide its urgency and route it to the right team.
Clef adds a vision encoder for image classification and a 64,000-token context window, both of which Cloudflare said are missing from Jev, the decision model that started the category when TypeSafe introduced it in September. Cloudflare said Clef reached a median latency of 209.3 milliseconds versus 524.1 milliseconds for Jev, about 2.5 times faster.
The models are released under the Apache 2.0 license and hosted on Hugging Face for local deployment. Alongside the release, Cloudflare said it is rolling out a reinforcement learning fine-tuning platform in two phases: hands-on help from its forward-deployed engineering team first, then a self-serve version that lets customers capture data, fine-tune and redeploy models on Cloudflare's infrastructure.
A decision model is a bet that most of what an agent does isn't writing prose, it's picking one of a few options fast and cheap. Shipping both the open weights and the fine-tuning pipeline to customize them is Cloudflare trying to own that infrastructure layer before it commoditizes.