Cloudflare releases Clef and Clef-Flash as open decision models

Cloudflare has released Clef and Clef-Flash, a pair of open decision models built to return bounded probability distributions instead of free-form text. Cloudflare’s October 1 announcement positions the models for workflow decisions where a system needs to choose among defined options quickly and consistently. The official model repositories describe Clef as a 27B multimodal model and Clef-Flash as a 9B variant. Both can consume state represented as text or JSON and can also use images or video before returning a probability for every allowed option in a typed question schema.

The artifact layer is unusually explicit. Cloudflare’s Hugging Face repositories are marked Apache-2.0, and the model cards describe standard sharded Safetensors for the backbone plus a separate joint schema head. Clef-Flash is post-trained from Qwen3.5-9B, while the larger Clef repository identifies Qwen3.8-27B as its backbone. The Clef-Flash card also states that its decision API is compatible with Jev and SystemOne. Benchmark tables published with the models remain Cloudflare-reported results and are not treated as independent measurements by Open Model Weights.

These releases are strong registry candidates, but the two variants should be verified independently rather than entered as one family-level record. Exact shards, joint-head files, context evidence, base-model lineage and license files need to be captured field by field. Hosted Workers AI availability and Cloudflare’s separate reinforcement-learning platform are ecosystem services; they should not be conflated with the terms attached to the downloadable model artifacts.