Cloudflare open-sources Clef decision models in 9B and 27B sizes

On October 21, 2026, InfoQ Chinese reported that Cloudflare had open-sourced Clef, a decision model released during its recent "Birthday Week" activities. The release includes two multimodal versions with 9 billion and 27 billion parameters, along with a platform for adapting them to specific decision tasks. According to Cloudflare product group manager Michelle Chen, principal machine learning engineer Alex Reneau, and senior engineering manager Kevin Flansburg, Clef takes a state and a schema of typed questions as input and can handle text, JSON, images, or video. In a single forward pass it returns probabilities for every available option of each question, without generating free text or parsing output. The results are meant to help AI agents choose next actions, such as routing customer support tickets, escalating them, or handing a decision to a human.

The report states that the smaller Clef-Flash has a median latency of 38.8 milliseconds in official benchmarks. Because the models are hosted on Cloudflare infrastructure, edge GPUs can reduce network latency and speed up decisions, allowing Clef to sit in the core path of an agent's decision loop and work with Workers AI to carry out follow-up actions. Its API is compatible with Typesafe AI's Jev System One model. The report names AI engineers, agent system architects, and SREs who need low-latency decision services as the intended audience. The available information stops at the open-source release and these technical details; the material does not mention future version plans or third-party evaluation results.