Cloudflare releases Clef open-weight decision models

Cloudflare has introduced Clef, a family of open-weight decision models, and announced an accompanying platform for reinforcement-learning fine-tuning. The company detailed both in a post on its engineering blog.

As the name suggests, Clef is oriented toward decision-making — the kind of model work where a system has to choose an action rather than simply generate free-form text. Releasing the weights openly gives developers the ability to inspect, run, and adapt the models on their own infrastructure, a notable move from a company best known for its edge network rather than for shipping foundation models.

Why it matters

The more interesting half of the announcement may be the tooling. Pairing open weights with a reinforcement-learning fine-tuning platform lowers the barrier for teams that want to specialize a model to their own reward signals and tasks, without assembling an RL pipeline from scratch.

• Open weights mean the Clef models can be self-hosted and modified.
• RL fine-tuning targets decision and action-selection use cases.
• Platform integration fits Cloudflare's push to host and serve AI workloads at the edge.

As an initial release, the full picture — model sizes, benchmarks, and licensing details — will come into focus as developers put Clef through its paces. For now, Cloudflare's entry signals growing interest in models built specifically for acting rather than just answering. Full details are available in Cloudflare's announcement.