Vibeleaderboard
← All Intel
Intel / article

Cloudflare ships Clef, open decision models for fast agent routing

Source
blog.cloudflare.com
Author
Cloudflare
Date

Context

Cloudflare's Clef and Clef-flash are decision models that accept the same API calls as TypeSafe's Jev, and Cloudflare says they are faster. A decision model does not write prose. A program gives it a description of the situation and a set of typed questions, and it returns an answer to each with a probability. In the launch post's example, the input says checkout has been failing for every customer for an hour. Clef is asked whether the request is urgent, which of three teams should handle it, and how severe the impact is on a four-step scale. Code can then route the ticket, escalate it, or hand it to a person based on those answers. Jev, which Simon Willison has described as returning numbers for categories and yes/no questions instead of text, drew a lot of to this kind of model, and Cloudflare built Clef so developers can switch to it easily. Cloudflare hosts both models on its Workers AI service and has published the weights on Hugging Face under the Apache 2.0 license.

The speed comes from how Clef produces an answer. Cloudflare took existing Qwen models, Qwen3.8-27B for Clef and the smaller Qwen3.5-9B for Clef-flash, kept their weights frozen, and trained a small added component that scores answers. At inference the model reads the input in one pass and then scores every valid choice in parallel, instead of generating text one token at a time and parsing an answer out of it. Across the 43 benchmarks Cloudflare ran, Clef's median response time was 209.3 milliseconds and Clef-flash's was 38.8, against 524.1 for Jev. Two other decision models in Cloudflare's table, Kev-9B at 51.4 and DiffusionGemma Jev at 84.4, were quicker than Clef but not Clef-flash. The one model faster than both Clef models, Laya at 5.8 milliseconds, scored far lower on Cloudflare's quality tests.

Cloudflare chose and ran all of these benchmarks itself, and the quality results are mixed. On the ten benchmarks it published, Clef beat Jev on eight but trailed on When2Call and BRIGHT. On TypeSafe's own workflow suite, Clef beat Jev on invoice processing, customer service and security incidents. Jev won on agent trace observability, and the customer service margin was 76.3 to 76.0. Cloudflare also says Clef can classify images, which it says Jev cannot yet do, and has a 64,000-token to Jev's 32,000. The speed advantage is large in Cloudflare's numbers. The quality lead is smaller and depends on the task, and all of these figures come from Cloudflare's own testing rather than an independent evaluation.

Terms in this piece · Glossary
  • attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
  • calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
Read the source blog.cloudflare.com
More from Cloudflare
Recommended reads
Comments

Checking sign-in…

Loading comments…