Cloudflare has announced Clef, a family of open-source decision models purpose-built for fast classification and agentic workflows, alongside a new reinforcement learning (RL) platform that lets developers fine-tune these models with their own data. The models are hosted on Workers AI, placing them directly in Cloudflare's edge inference stack so applications can make routing, labeling, and decision calls close to users.
- Clef and Clef-flash: open-source decision models designed for high-speed classification and agentic workflows.
- Hosting on Workers AI, making them callable from within the Cloudflare developer platform.
- A new reinforcement learning platform for fine-tuning decision models using your own data.
- Open-source availability, so teams can inspect, adapt, and self-host the models.
Clef targets the kind of lightweight, repeatable decisions that agentic systems make constantly: classifying an input, choosing a branch, or gating the next action. Rather than prompting a large general-purpose LLM for each of these steps, developers can call a smaller decision model tuned for the task. The Clef-flash variant emphasizes latency, which matters when a model sits inside a request path or an agent loop where every millisecond compounds. Running on Workers AI keeps inference inside Cloudflare's network, reducing the round trips that typically accompany external model APIs.
The RL fine-tuning platform is the more consequential piece for teams with proprietary data. Reinforcement learning lets a model improve against a reward signal rather than only imitating labeled examples, which suits decision tasks where correctness is easier to score than to demonstrate. Developers can supply their own data to specialize a base Clef model for domain-specific classification or agent policies.
For developers, the appeal is a tighter loop: open weights to audit, edge inference to reduce latency, and a fine-tuning path that turns production data into better decisions. Agent builders evaluating cost and speed per decision now have a self-hostable alternative to general-purpose models.