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

Cloudflare introduces Clef and Clef-flash, two decision models trained by Cloudflare and hosted on Workers AI, designed to produce structured, deterministic outputs for workflow decisions. These models are optimized for speed and accuracy, outperforming general LLMs in latency while maintaining compatibility with the Jev Decision Index benchmark. Clef currently leads the benchmark, offering faster classification with typed outputs that agents can use programmatically.

The models are fully open-sourced under an Apache 2.0 license on Hugging Face, allowing users to run them locally or integrate them into their systems. Cloudflare also debuts a new reinforcement learning product enabling customers to fine-tune Clef for their specific use cases, enhancing adaptability for agentic workflows.

Cloudflare’s Clef models include a vision encoder for image classification and feature a 64k context window, surpassing competitors like Jev’s 32k window. Benchmarks show Clef and Clef-flash outperform other decision models in latency and accuracy, with Clef-flash excelling in speed-critical scenarios.

Hosted on Cloudflare’s edge infrastructure, Clef models benefit from low network latency and strict data privacy guarantees. The larger Clef model prioritizes precision, while Clef-flash is optimized for latency-sensitive decisions, both supporting enterprise workloads without storing or training on user data unless fine-tuning is explicitly enabled.