Cloudflare launches Clef decision models for AI agents

Cloudflare has introduced Clef and Clef-flash, two specialized decision models designed to help AI agents make autonomous choices. Unlike traditional large language models that generate text, these models evaluate inputs and assign probabilities to predefined options. This approach allows downstream software to instantly trigger actions, such as routing customer support tickets or escalating security threats. According to Cloudflare, the technology means "a human does not necessarily need to be in the loop" for routine agentic decisions.

The new models aim to outperform TypeSafe AI's rival model, Jev, primarily on speed and capability. Built on the Qwen3.8-27B architecture, Clef features a 64,000-token context window, which is double the capacity of Jev. The smaller Clef-flash is based on Qwen3.5-9B. While Jev is currently limited to text, both Clef models can process both text and images. In benchmark tests, Clef-flash achieved a median latency of 39 milliseconds, and Clef recorded 209 milliseconds. Both significantly beat Jev's median response time of 524 milliseconds.

Cloudflare is already utilizing Clef internally for threat intelligence. In one test, the model fetched, rendered, and classified a website in 2.2 seconds, whereas the company's fastest general-purpose language model required 4.7 seconds. To train the models, Cloudflare kept the base Qwen models intact and used synthetic data alongside its own version of Reinforcement Learning for Calibrated Decisions to train auxiliary components.

Both models are now available on Cloudflare's Workers AI platform and Hugging Face under an Apache-2.0 license. Cloudflare also plans to roll out a reinforcement learning service to help clients fine-tune Clef for specific tasks. This service will leverage containerized sandboxes and technology from Replicate, a company Cloudflare acquired in late 2025, to deploy customized models directly to the edge.