Cloudflare has released Clef and Clef‑flash, two open‑weight decision models hosted on Workers AI and published under an Apache 2.0 license on Hugging Face. Decision models output bounded, structured results—such as probabilities for predefined choices—so agents can make programmatic decisions without the open‑ended text generation of large language models. Clef and Clef‑flash are fully compatible with Typesafe’s Jev API, allowing them to drop into existing workflows for tasks like routing support tickets or classifying threat‑intelligence data. Unlike Jev, they add a vision encoder that can accept image inputs and expand the context window to 64 k tokens, more than double Jev’s 32 k limit.
On decision‑making benchmarks from the Jev Decision Index, Clef and Clef‑flash score competitively, often exceeding Jev. For example, Clef achieves 98.47 % on BFCL case‑exact, 69.19 % on ToolRet nDCG@10, and 91.93 % on API‑Bank accuracy, while Clef‑flash reaches 98.76 %, 66.43 %, and 93.11 % respectively. Latency improvements are striking: Clef‑flash returns a result in a median of 38.8 ms, compared with 524.1 ms for Jev, and the larger Clef model averages 209.3 ms. These gains, combined with edge‑GPU deployment on Workers AI, let developers place Clef in the hot path of agentic systems for fast, reliable decisions.
The image illustrates the trade-off between decision-making accuracy and response time, showing Clef-flash's superior latency and competitive benchmark results.
Beyond the base models, Cloudflare offers a reinforcement‑learning fine‑tuning service. Customers can work with Cloudflare’s forward‑deployed engineer team to build datasets via AI Gateway, generate rollouts on Workers AI, score actions in Container‑based RL sandboxes, update weights with a dedicated Trainer, and redeploy the customized model using Workers AI’s Bring Your Own Model capability. The training starts from Qwen backbones (Qwen3.8‑27B for Clef, Qwen3.5‑9B for Clef‑flash), applies label‑smoothed cross‑entropy and Brier loss for calibrated probabilities, and uses a novel Reinforcement Learning for Calibrated Decisions (RLCD) objective to improve accuracy while preventing distribution shift. This lets organizations adapt Clef to specialized workloads such as Trust & Safety classification or bot detection without sacrificing the speed and structure that decision models provide.