CLOUDFLARE has released two open-source AI decision models, Clef and Clef-flash, designed to speed up AI agents by making rapid routing and classification decisions on-the-fly. Integrated with Cloudflare’s Workers AI platform, the models are also available with their weights on Hugging Face under the Apache 2.0 licence. Built to compete with the Jev decision model from TypeSafe, Clef and Clef-flash are designed for fast, text- and image-informed judgments rather than free-form generation.
The Clef family eschews traditional autoregressive text generation in favour of pre-defined options and scoring criteria. After receiving an input state, Clef uses cross-attention to evaluate all options in parallel and returns a probability score for each one. Benchmarks across 43 tasks show notable speed gains: Clef-flash delivers a median latency around 38.8 milliseconds, Clef about 209.3 milliseconds, versus Jev at 524.1 milliseconds.
Clef is built on the 27‑billion-parameter Qwen3.8-27B, while Clef-flash uses the 9‑billion-parameter Qwen3.5-9B, both retaining Qwen’s vision encoder, allowing input of text, JSON and up to four images. The API is drop-in compatible with Jev’s System One specification, enabling developers to migrate existing Jev apps with minimal friction.
Cloudflare also plans an RL fine-tuning service, initially via its in-house FDE team for enterprise customers, with a future self-service platform for data preparation and deployment.