Introducing Clef: our open-source decision models, and new RL fine-tuning platform

Cloudflare today released Clef and Clef-flash, two Cloudflare-trained decision models hosted on Workers AI, and debuted a reinforcement learning product that lets customers fine-tune Clef for their own workloads.

A decision model returns bounded, typed answers with probabilities rather than open-ended text. Cloudflare tested Clef on its Threat Intelligence team to classify website domains with Browser Run: one classification took 2.2s to fetch, render and classify, versus 4.7s for gpt-oss-120b in the same workflow, which returned only two classifications. Clef has a vision encoder for image classification, which Cloudflare says Jev lacks, and a 64k context window versus Jev's 32k.

Cloudflare says Clef currently leads the Jev Decision Index, and that across 43 eval benchmarks its models beat competing decision models on latency except Laya. Median latency is 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev. On Typesafe's own eval suite, Clef models beat Jev in 3 of 4 areas. Both models are Jev-API compatible and produce strictly typed outputs.

Clef is built on a frozen Qwen3.8-27B backbone and Clef-flash on Qwen3.5-9B, with a routing head jointly optimized alongside rank-256 low-rank adapters. Inference uses a Qwen prefill-only pass, then scores valid schema choices in parallel; the decision step is non-autoregressive. Post-training used label-smoothed cross-entropy plus Brier loss, and a method Cloudflare calls Reinforcement Learning for Calibrated Decisions (RLCD) that grants partial credit to adjacent ordinal choices and applies a reference penalty.

The models are open-sourced on Hugging Face under Apache 2.0. Cloudflare guarantees it does not read, store or train on requests or responses unless the fine-tuning product is used. Fine-tuning is offered first through Cloudflare's forward-deployed engineer team, with a self-serve platform to follow; its RL service combines AI Gateway, Workers AI, Containers and a new Trainer component.

Why it matters: developers can run Clef locally, swap it into Jev-based workflows through API compatibility, and put low-latency decisions in an agent's hot path alongside Workers AI LLMs. Enterprise users get a no-training-on-data guarantee unless they opt into fine-tuning.