Server racks in a data center — Photo by panumas nikhomkhai / Pexels

Cloudflare ships Clef: open-weight decision models on Workers AI

Cloudflare’s first in-house models — Clef (27B) and Clef-flash (9B) — land on Workers AI under Apache 2.0 with free output tokens.

THE HOOK

Cloudflare just trained its own models — and they’re decision engines, not chatbots.

THE GIST

Cloudflare released Clef (27B, Qwen3.8) and Clef-flash (9B, Qwen3.5) on Workers AI under Apache 2.0 on Hugging Face — Jev-compatible decision models for agentic routing. Input is $0.24/$0.09 per million tokens with free outputs, plus a new RL fine-tuning platform.

THE TAKE

This is edge infra flexing into model ownership. Prefill-only scoring beats token-by-token chat for routing. Free output tokens is a knife fight with hosted classifiers. If your agents need yes/no/which-team in under 50ms, chat GPUs were always the wrong tool.

OPINION

Birthday Week dumps can look like noise, but open weights plus a hosted edge path is a real wedge against Typesafe’s Jev. The Clef announcement and Hugging Face drops make the experiment cheap — the RL fine-tune loop is where Cloudflare wants you locked in.

  • Story type: Breaking news
  • Impact: Medium
  • Confidence: High (official announcement)
  • Audience: Developers
  • What changed: Cloudflare’s first in-house open-weight decision models live on Workers AI with free output tokens
  • What to watch: Whether Clef displaces Jev in production agent routers and how the RL fine-tune product prices
  • Reading time: 2 minutes