Cloudflare Releases Clef Decision Models
Cloudflare released 27B Clef and 9B Clef-flash, open-weight multimodal decision models that return probabilities for typed questions instead of generating free-form text. They run on Workers AI, support Jev-compatible APIs, and come with an RL fine-tuning platform.
Cloudflare is turning decision models into an infrastructure feature for agent workflows, pairing fast structured inference with edge deployment and customization.
- –Clef and Clef-flash handle text, JSON, images, and video, producing schema-bound probabilities without output parsing.
- –Clef-flash reportedly reaches 38.8 ms median latency, making it practical for routing, triage, and tool-selection loops.
- –Jev compatibility lowers migration costs for developers already building against the emerging decision-model interface.
- –The Apache 2.0 weights are open, but Cloudflare has not released its training data or full pipeline, making this open-weight rather than fully reproducible open source.
- –Cloudflare’s benchmark claims are promising but vendor-run, so independent evaluations will determine whether the accuracy advantage holds.
DISCOVERED
1h ago
2026-10-01
PUBLISHED
1h ago
2026-10-01
RELEVANCE
AUTHOR
steipete