Clef
Cloudflare's open-weight 27B multimodal decision model, post-trained from Qwen3.8-27B. A joint schema head scores every option of every noul, choice and score question in one prefill-only pass over text, JSON, images or video. Also hosted on Workers AI.
Clef returns one logit per allowed option (softmax per question); nothing is generated. It needs the repo's own joint_schema_model.py, which also exposes a Jev/SystemOne-style systemone() call. The hosted context is 65,536 tokens; the local code defaults to 16,384. Cloudflare's blog says training used a Brier loss, but no calibration figure is published. Sibling: Clef-flash (Cloudflare/clef-flash, 9B on Qwen3.5-9B; 38.8 ms median and 122.4 ms p95 request latency in Cloudflare's run; $0.09 per million input tokens on Workers AI). Cloudflare's own run of the community Decision Index 0.2.1 suite (a third-party board) gives Clef 61.21 and Clef-flash 57.07 on 36 of 38 panel benchmarks; these are self-reported, not the board's own measurement. Third-party quantisations exist and are not the maker's.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 27B |
| Base model | Qwen/Qwen3.8-27B |
| Maker | Cloudflare |
| Released | 2026-09-30 |
| License | apache-2.0 |
| Reported latency | 209.3 ms p50 / 238.6 ms p95 request latency (hardware not stated; the card was tested on one H200) |
Hosted API
Served by Cloudflare Workers AI — $0.24/MTok input. Get access · API docs.
Get the weights
pip install systemonemodels
systemone pull cloudflare/clef
The files are served from the maker's Hugging Face repository, Cloudflare/clef, and verified against the checksums recorded here.