Cloudflare launches open-weight Clef decision models for AI workflows

Cloudflare has introduced Clef and Clef-flash, two open-weight AI models designed to make fast, structured decisions inside automated workflows. The company positions them as alternatives to Typesafe AI’s Jev, a decision model built to answer bounded questions such as yes or no, multiple choice, and ranked options.

The models are available through Cloudflare Workers AI and as downloadable weights under an Apache 2.0 license. In its official blog post, Cloudflare says Clef can classify information, return typed results with probabilities, and help AI agents decide when to route, escalate, act, or involve a human.

Decision models differ from general-purpose chatbots. A large language model can write text, summarize documents, and use tools, but its answers can vary between runs. Decision models target narrower tasks. They are intended to return a constrained answer in a predefined format, such as whether a support request is urgent, which department should receive it, or how severe an incident is.

Cloudflare says it has tested Clef in its threat intelligence work. In one example, the company used the model with a browser tool to inspect and categorize websites. Clef classified a website in 2.2 seconds, according to Cloudflare, compared with 4.7 seconds for the company’s gpt-oss-120b model in the same workflow. Cloudflare says Clef also returned more categories in that test. These results are company-reported and have not been independently verified in the supplied sources.

Faster decisions, according to Cloudflare’s tests

Clef is based on a post-trained, frozen Qwen3.8-27B model. Clef-flash uses the smaller Qwen3.5-9B model. Rather than generating a response token by token, the models assess permitted answer choices in parallel after processing the input. Cloudflare says this design reduces response time while keeping outputs within a specified schema.

On Cloudflare’s reported benchmark results, Clef recorded a median latency of 209.3 milliseconds and Clef-flash 38.8 milliseconds. Jev recorded 524.1 milliseconds in the same comparison. The company says its models outperformed Jev in many tests involving tool selection, customer service classification, banking queries, and structured API tasks. However, Jev scored higher in some areas, including the When2Call and BRIGHT benchmarks.

Cloudflare also says Clef supports images and has a 64,000-token context window. The Register reports that the model can handle images and video, and notes that Jev can process up to 64,000 tokens in a request, although its state plus longest individual question is limited to 32,000 tokens.

The company says Clef is compatible with Jev’s API, which could let existing Jev users test the new models without redesigning their applications. Cloudflare describes Clef as the model for higher-precision work and Clef-flash as the option for latency-sensitive tasks.

Open weights, but not open training data

Cloudflare calls the models open source, but their training datasets are not public. Brandon Vigliarolo reports for The Register that Clef-flash requires at least 41 GB of GPU memory for local operation, while Clef requires 85 GB under Cloudflare’s stated assumptions of single concurrency and a 64,000-token context window.

The hosted service costs $0.24 per million tokens, according to The Register. That is substantially above the $0.042 per million tokens cited for Jev. Cloudflare argues that edge hosting through Workers AI can reduce network delays for customers that place decisions directly in an automated process.

Cloudflare is also offering hands-on fine-tuning services through its forward-deployed engineering team. It plans a self-service product that would let customers collect workflow data, train a customized Clef model, and deploy it through Workers AI. The company sees applications in support triage, trust and safety reviews, and bot classification.

Sources

• Clef: our open-source decision models – Cloudflare Blog
• Cloudflare debuts open-weight multimodal decision models Clef and Clef-flash, claiming they are smarter and faster than Jev, based on Qwen3.8-27B and Qwen3.5-9B – The Register