Clef (27 billion parameters) and Clef-flash (9 billion) do not generate text: they return typed answers with probabilities, in open source under the Apache 2.0 license.
On October 1st, Cloudflare unveiled Clef and Clef-flash, the first models trained by its own teams and hosted on Workers AI. They belong to the category of "decision models", popularized in recent weeks by Typesafe's Jev: instead of writing a response, they read an input status and a list of typed questions, then assign a probability to each authorized response. An agent can thus know if a support ticket is urgent and which team to send it to, without going through text generation.
More context, vision and speed
Clef is based on Qwen3.8-27B and Clef-flash on Qwen3.5-9B, frozen and supplemented by low-rank adapters and a routing head trained to calibrate the probabilities. The decision is made in a single pass, without generating token by token. Unlike Jev, which is limited to text, both models also accept images and video, with a context of 64,000 tokens compared to 32,000. Out of 43 evaluations, Cloudflare announces a median latency 2.5 times lower than Jev for Clef and 13 times for Clef-flash, and first place in seven of the ten decision tests selected.
The weights are published on Hugging Face under the Apache 2.0 license and hosted usage is billed at $0.24 per million input tokens for Clef, $0.09 for Clef-flash. Cloudflare is also launching a fine-tuning offer using reinforcement learning, first supported by its engineers, then in self-service. The company says it is already using Clef to classify domain names within its threat intelligence team.
Read the full article on the original source
Go to the source →