Cloudflare has released decision models, including open-weights Clef
Cloudflare has released its own decision models, including Clef, an open-weights model free to download on Hugging Face.
Cloudflare has released its own decision models, Clef and the smaller Clef-flash, hosted on Workers AI. It has also made the Clef weights free to download on Hugging Face under an Apache 2.0 license. The company's blog post, dated October 1, 2026, presents the pair as Jev-compatible alternatives to TypeSafe AI's model. The startup launched Jev in September, and it has drawn a wave of imitators.
A decision model takes a body of input, such as a support message, an email, or a web page, and answers a fixed set of questions. Each answer is one of a predefined set of options, returned with a probability. Cloudflare's blog gives a support-ticket example. The model returns whether a message is urgent and which team should handle it, and the calling code uses those typed answers to route the ticket or escalate it to a person. The blog says Clef adds two things Jev lacks. It has a vision encoder that can classify images, which Jev does not do, since Jev handles text only. It also has a 64k-token context window, compared with Jev's 32k.
The mechanism explains the speed claim. A standard LLM writes its answer one token at a time, with each new token depending on everything before it. Michelle Chen, a group product manager at Cloudflare, described the difference: "Our decision models don't do that." According to the blog, Clef runs a prefill-only pass over the input, then scores every valid schema choice in parallel, so no intermediate text is generated. The Hugging Face model card describes Clef as a 27B multimodal model built on Qwen/Qwen3.8-27B, with a small joint head that outputs one logit per allowed option for each question. A softmax over those logits gives the probabilities. The card says there is no free-form text generation and no output parsing.
That same constraint is why TypeSafe AI can say Jev cannot hallucinate. The model can only pick from options it was given, so it cannot produce an invented answer. Carnegie Mellon's Ramayya Krishnan cautioned, though, that "eliminating hallucinations in the output does not eliminate errors in decision-making." In his group's tests of Jev, the model did well when the answer could be read directly from the prompt. Accuracy fell on questions that needed math or a logic puzzle.
The latency numbers come from Cloudflare's own runs. The blog reports a median of 209.3 ms for Clef, 38.8 ms for Clef-flash, and 524.1 ms for Jev. The 95th-percentile figures are 238.6, 122.4, and 536.0 ms. Chen gave a more practical example from Cloudflare's threat-intelligence team, which uses Clef to classify website domains by category and phishing risk. That task took 2.2 seconds to fetch, render and classify a site. gpt-oss-120b, which Cloudflare called its fastest general LLM, took 4.7 seconds and returned only two classifications. The comparison is with an LLM rather than with Jev.
On the Decision Index, Clef scores 98.5 on BFCL case exact accuracy, against Jev's 95.8, and 97.4 on CLINC150+OOS macro-F1, against Jev's 89.3. Jev leads on several general-knowledge and reasoning tests, though. MMLU-Pro is 65.9 for Clef against 82.7 for Jev, BBH is 73.7 against 92.9, and GPQA Diamond is 48.0 against 78.3. On Cloudflare's four-workflow evaluation drawn from TypeSafe Evals, the blog says Clef beat Jev in three areas and trailed in agent trace observability, at 68.5 against 71.6. The Hugging Face card describes the Decision Index results as coming from Cloudflare's internal run of its own suite.
Pricing is the other lever. Cloudflare lists Clef at $0.24 per million input tokens on Workers AI, with no charge for output tokens. Images are converted to input tokens, capped at 1,024 tokens each. TypeSafe AI charges US $0.042 per million input tokens, and its output is free, so Jev is cheaper per token. The Workers AI page also lists a context window of 65,536 tokens. The blog adds that Cloudflare does not read, store or train on requests to its hosted models, unless a customer opts into fine-tuning.
Clef has an earlier lineage. The blog says Cloudflare's first decision-model work adapted the DiffusionGemma model to expose the probabilities of its outputs. That initial approach built on independent research by Matt Mastracci. Clef keeps the idea but uses Qwen as its base. Cloudflare has also announced a reinforcement-learning fine-tuning product, offered first through its forward-deployed engineers and later as a self-serve platform.
Krishnan's group also tested a hybrid setup, in which Jev's low-confidence answers went to GPT-6 for a second look. The combination slightly beat GPT-6 alone and cost 41 percent as much. Zhongyu Wei, a Fudan University professor who co-authored a benchmark of decision models, expects that design to spread. "The decision model can call an LLM to give a more detailed explanation. And an LLM can call a decision model for a quick judgment. I think that's very flexible."
What changed
Cloudflare released its own decision models, Clef and the smaller Clef-flash, hosted on Workers AI, and made the Clef weights free to download on Hugging Face under an Apache 2.0 license.
Type: New capability · Applies to: Clef, Clef-flash, Workers AI
Why it matters
Open-weights decision models let teams run classification locally without fine-tuning a separate classifier for each task. Companies using Cloudflare services may also see the same approach in their own security tooling.
What you can do
Evaluate open-weights decision models such as Clef for triage tasks like spam, phishing or prompt-injection detection.
How the story developed
Adds that Cloudflare's AI Platform lead describes Clairic as a separate extraction model, says Clef was built in about two weeks, and says Cloudflare plans post-training work and possibly reasoning or other LLMs.
Original source
Further reporting
Where we saw it
IEEE Spectrum, Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works