Cloudflare Open Sources Decision Models for AI Agents

During its recent Birthday Week, Cloudflare announced Clef, a set of open-weight AI models built to choose between predefined options rather than generate text. Cloudflare released 9B- and 27B-parameter models along with a platform for adapting them to specific decision-making tasks.

Clef is a 27B multimodal model that takes a state and a schema of typed questions as input and returns decisions. It can process text, JSON, images, or video, and returns a probability for each allowed option for every question in a single forward pass, without generating free-form text or requiring output parsing. Cloudflare's API is compatible with the Typesafe AI Jev System One model. A decision model classifies inputs and returns typed outcomes with probabilities, allowing AI agents to use the results to determine how to act, such as routing a support request, escalating it, or deferring to a human.

Cloudflare also released Clef-Flash, a smaller 9B multimodal model designed for latency-sensitive decisions, with a median latency of 38.8 ms in Cloudflare's benchmarks versus 209.3 ms for the 27B Clef model. According to Michelle Chen, Alex Reneau, and Kevin Flansburg, Clef differs from other decision models by including a vision encoder that classifies visual content, unlike Jev, which does only text classification today, and by having a 64k context window compared with Jev's 32k.

A Hacker News thread debated the category. Jacek Złydach called it a low-hanging fruit rather than a new paradigm, crediting Typesafe as first to pick it up and market it, while user SebastianSosa warned that public benchmarks are easy to cheat and could distract from useful work. On Reddit, bugra_sa said knowing when Clef should punt matters more than its raw accuracy score.

Cloudflare plans to fine-tune Clef for use cases such as support triage and bot classification using its historical labelled data. It announced a fine-tuning service for customers to adapt Clef to their own workloads, initially with support from Cloudflare engineers, with a self-service platform planned later; no firm date has been announced. The models are available through Workers AI and as downloadable weights on Hugging Face.

Why it matters: Developers can run Clef decisions on Cloudflare's edge GPUs for low network latency, potentially placing it in the hot path for agents alongside an LLM on Workers AI, and can experiment with the weights or request fine-tuning for their own workloads.