Clef and Clef-flash

Cloudflare released Clef and Clef-flash, open-weight decision models under Apache 2.0 that return probabilities for typed answers, hosted on Workers AI and Hugging Face.

These are the first models trained by the Workers AI team. They answer yes/no, choice and score questions with probabilities instead of text. Clef is post-trained from Qwen3.8-27B and Clef-flash from Qwen3.5-9B, with a non-autoregressive decision step after one prefill pass. They accept images, have a 64k context window and are compatible with the Jev API. Cloudflare also announced an RL fine-tuning product for Clef.

Date
Thursday 1 October 2026
Lab
Cloudflare
Kind
open-weights
Access
open weights
Price
$0.24 per million tokens for Clef on Workers AI

Figures

MeasureValueMeasured by
BFCL case exact98.47 (Clef), 98.76 (Clef-flash), 95.75 (Jev)
Jev Decision Index, self-reported
company
Median latency209.3 ms (Clef), 38.8 ms (Clef-flash), 524.1 ms (Jev)
43 eval benchmarks
company
Context window64k tokens
The Register says Jev also handles 64k per request, but its state plus longest question is limited to 32k
company
Price per million tokens$0.24 (Clef)
The Register says Jev is $0.042
The Register

Benchmark scores are self-reported and have not been reproduced on the official Jev Decision Index.

Sources

  1. blog.cloudflare.com/clef-decision-models/
  2. theregister.com/ai-and-ml/2026/10/01/cloudflare-tries-to-outplay-jev-with-open-weight-clef
  3. marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/

This record was checked and corrected against its sources on 8 October 2026. How we check

Read the daily brief for 1 October 2026