- Published on
Cloudflare Open Sources Decision Models for AI Agents
- Authors
- Name
- DeepHorus
Cloudflare has open-sourced Clef, a family of open-weight AI models designed to make decisions rather than generate text. The announcement was made during the company's recent "Birthday Week" event. Cloudflare released 9B- and 27B-parameter models, along with a platform for adapting them to specific decision-making tasks. Instead of producing free-form language, Clef selects between predefined options and returns probabilities for each allowed choice. The move aims to give AI agents a more reliable way to act.
Decision models like Clef classify inputs and return typed outcomes with probabilities. AI agents can use these results to determine how to act, such as routing a support request, escalating it, or deferring the decision to a human. Cloudflare's API is compatible with the popular Typesafe AI's Jev System One model, which also focuses on decision-making. This compatibility allows developers to compare or switch between the two systems. The approach eliminates the need for output parsing and reduces ambiguity in agent workflows.
The flagship 27B Clef model is multimodal, accepting text, JSON, images, or video. It takes a state and a schema of typed questions as input and returns a probability for every allowed option in a single forward pass. A smaller 9B model, Clef-Flash, targets latency-sensitive decisions. In Cloudflare's benchmarks, Clef-Flash achieves a median latency of 38.8 ms, compared with 209.3 ms for the 27B model. Clef also features a 64k context window, double Jev's 32k, and includes a vision encoder for visual classification.
Michelle Chen, group product manager, Alex Reneau, principal machine learning engineer, and Kevin Flansburg, senior engineering manager at Cloudflare, explained the edge-hosting advantage. "Because they are hosted on Cloudflare's infrastructure, we're able to take advantage of our GPUs at the edge, leading to low network latency and faster decisions," they wrote. This means Clef can sit in the hot path for agents, combined with an LLM on Workers AI to take action. Cloudflare plans to fine-tune Clef for support triage and bot classification using historical labelled data. Such tuning could improve both accuracy and speed for enterprise customers.
The release has sparked debate on Hacker News about whether decision models represent a new category. Jacek Złydach argued it is "not a 'new paradigm,' it's a low-hanging fruit that's been lying around for years." Typesafe, he added, was first to "stop and pick it up, and market the shit out of it." Another user, SebastianSosa, warned that public benchmarks are easy to cheat and could distract developers from building useful systems. On Reddit, user bugra_sa said he cares more about whether Clef knows when to punt than its raw accuracy score.
Cloudflare also announced a fine-tuning service that lets customers adapt Clef to their own workloads using their data. Initially, Cloudflare engineers will provide support, with a self-service platform planned for a later release. No firm date has been announced for that platform. The company has made the Clef models available through Workers AI and as downloadable weights on Hugging Face. Developers are invited to experiment with the models and provide feedback. This open approach could accelerate adoption and refine the models for real-world decision tasks.
