Clef
Cloudflare's open-weight decision model — typed probabilities instead of prose, 27B, Apache 2.0.
About
Clef, released October 1, 2026 during Cloudflare's Birthday Week, is the company's first model trained by its own Workers AI team — and it never writes a sentence. It is a decision model: you send a state description plus a schema of typed questions (yes/no, multiple choice, scored rubrics) and it returns a calibrated probability for each allowed option, with up to 64 questions in a single request.
Clef has 27 billion parameters built on a Qwen3.8-27B backbone: Cloudflare froze the backbone and trained low-rank adapters on top. It carries a 64K-token context on Workers AI and reads text, JSON, images, and video frames. It is API-compatible with TypeSafe's Jev (the System One format), so migrating is mostly an endpoint change. Cloudflare reports a median latency of 209.3 ms — about 2.5x faster than Jev's 524.1 ms in its own tests — and bills $0.24 per million input tokens on Workers AI with output free.
Weights ship under Apache 2.0 on Hugging Face, so teams can self-host and fine-tune. Cloudflare also announced a reinforcement-learning tuning service for private data, starting with forward-deployed engineers.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: median 209.3 ms vs Jev 524.1 ms (~2.5x). Independent third-party bench (Oct 10, 2026): 71/82 correct, median 0.517 s over the network.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Release facts (October 1, 2026; 27B on Qwen3.8-27B; 64K context; Apache 2.0 weights on HF; Workers AI availability) corroborated across Cloudflare's announcement and multiple independent deep-dives with consistent figures. Docked for: (1) the latency and benchmark comparisons are Cloudflare's own with no independent rerun at release; (2) the RL fine-tuning service is engineer-led, not self-serve; (3) as a decision model it cannot explain or handle anything outside its question schema.