Atlas AI Resources

← Back to AI News

AI products

Cloudflare Released Clef Decision Models. They Return Probabilities, Not Essays.

Published October 2, 2026

On October 1, Cloudflare announced Clef and Clef-flash, the first models trained by its Workers AI team. They are not chatbots. You give them a situation and a short list of structured questions—yes/no, pick one option, or rate on a scale—and they return a probability for every allowed answer. The models run on Workers AI, and the weights are open under Apache 2.0 on Hugging Face.

What happened

In a Cloudflare blog post by Michelle Chen, Alex Reneau, and Kevin Flansburg, the company described a decision model: software that chooses among fixed options instead of writing free-form text. Picture a support ticket. You ask whether it is urgent, which team should own it, and how severe the impact is. The model replies with numbers your code can act on—route, escalate, or wait—without parsing a paragraph of prose.

Clef is the larger model, built on a frozen Qwen3.8-27B backbone. Clef-flash is the smaller, faster sibling on Qwen3.5-9B. Both keep a vision encoder, so they can take images (up to four per request on Workers AI) as well as text or JSON. Context is listed at 64K tokens. The API matches TypeSafe AI’s Jev “System One” style: change the endpoint and model name if you already use Jev. Cloudflare also said it is starting a reinforcement-learning fine-tuning service, first with forward-deployed engineers, then as self-serve.

Cloudflare reports its own benchmark numbers. Across a shortlist of ten Decision Index tests, a Clef model scored highest on seven. Examples it highlights include BANKING77 macro-F1 of 94.20 for Clef versus 79.74 for Jev, and median latency of 209.3 ms for Clef and 38.8 ms for Clef-flash versus 524.1 ms for Jev in that internal run. Hosted input pricing on Workers AI is listed at $0.24 per million tokens for Clef and $0.09 for Clef-flash. Those scores, timings, and prices are vendor-reported.

The Register spoke with Chen. She confirmed the training datasets are not public, even though the weights are Apache 2.0, and said self-hosting needs roughly 85 GB of GPU memory for Clef and 41 GB for Clef-flash at single concurrency with a full 64K context. The Register also noted that Cloudflare’s Decision Index scores have not yet been independently reproduced on the official leaderboard.

Why it matters to a normal person

Most people meet AI as a chat box that writes. Decision models sit elsewhere in the stack: the moment software has to choose—spam or not, which queue, block or allow—without waiting for a long generated answer. If agents (programs that take multi-step actions) become common, those tiny, typed choices show up constantly. A faster, more predictable answer at that step can mean fewer wrong escalations and less “the bot wrote something weird and we had to parse it.”

Open weights also matter for shops that cannot send every support ticket or security event to a third-party API. You still need serious GPU hardware to run Clef locally, per Cloudflare’s own guidance. For everyone else, Workers AI is the ready path. Competition with Jev and smaller open models is already about vision, context length, latency, price, and who you trust with the data.

What it does not mean

It does not mean Clef replaces general chat models for writing, coding, or open-ended reasoning. It answers only the questions you define. It does not mean independent labs have confirmed Cloudflare’s leaderboard—MarkTechPost and The Register both flag that the published scores are Cloudflare’s own runs so far. And “open source” in the launch language means open weights under Apache 2.0, not a public training set.

It also does not mean every product can drop humans out of the loop. Probabilities still need thresholds, logging, and a plan for when the model is unsure. A high score on a bank-intent benchmark is not a guarantee on your private tickets.

Sources