Cloudflare released two open-weight decision models, Clef and Clef-flash, on October 1. They read a situation and a set of typed questions and return decisions, instead of writing text. The weights are on Hugging Face under the Apache 2.0 licence, and both models run on Cloudflare’s Workers AI service, which the company says can stand in for Jev’s.
New to this? Read it in simple words
- Cloudflare released two free AI models, Clef and Clef-flash, that make choices for AI agents.
- They look at a situation and pick answers, and they can also read images.
- Cloudflare says they beat the paid model Jev on most of its tests and are faster.
- No outside group has checked these results yet.
- Decision model
- An AI model that only chooses between set options, rather than writing answers.
- Open weights
- The trained model is published, so anyone can download and run it.
- Multimodal
- Able to work with more than one kind of input, such as text and images.
What Cloudflare released
Clef has 27 billion parameters and is post-trained from Alibaba’s Qwen3.8-27B. Clef-flash has 9 billion and is built on Qwen3.5-9B. Both can read up to about 64,000 tokens.
The model card calls Clef “a 27B multimodal model that turns a state and a schema of typed questions into decisions”. It can look at images as well as text.
On Workers AI, Cloudflare charges $0.24 per million input tokens for Clef and $0.09 for Clef-flash, and nothing for output, MarkTechPost and The Register report.
Cloudflare’s own measurements, from its blog and changelog. No outside test yet.
Cloudflare’s own numbers
“Across 10 decision benchmarks, a Clef model scores highest on 7,” Cloudflare’s changelog says. On one banking test, it reports 94.20 for Clef against 79.74 for Jev.
Cloudflare measured a median time of 209.3 milliseconds per decision for Clef, 38.8 for Clef-flash and 524.1 for Jev.
Jev stays ahead on general reasoning tests in the same tables, such as GPQA Diamond, at 78.3 against 48.0 for Clef.
Why it matters
A day after Amazon’s Strands Decider, Clef gives teams a second free model they can run on their own machines instead of renting Jev.
Cloudflare also opened a reinforcement-learning service to fine-tune the models. At first its engineers run it with customers; a self-serve version comes later, it says.
All numbers are vendor-reported, with no independent replication yet, MarkTechPost notes. The Register says checks on an official decision index are still pending.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 1, 2026Introducing Clef: our open-source decision models, and new RL fine-tuning platform Cloudflare
- 2
- 3
- 4Research · October 1, 2026Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text MarkTechPost
- 5Research · October 2, 2026Cloudflare Open Sources Clef to Challenge OpenAI and Amazon on AI Agent Decisions Startup Fortune
We used Cloudflare’s blog, its changelog and the model card, and compared The Register, MarkTechPost and Startup Fortune. All scores and speeds come from Cloudflare’s internal runs. Outlets describe the benchmark wins in different ways, so we quote Cloudflare’s changelog. We left out memory figures that only one outlet gives.