Cloudflare Open Sources Clef Decision Models for AI Agents

Cloudflare Introduces Clef Decision Models

Cloudflare announced Clef, a new set of open-weight AI models during its "Birthday Week." These models, available in 9B and 27B parameter versions, are designed to select from predefined options rather than generating free-form text. Cloudflare also provided a platform for adapting these models to specific decision-making tasks.

Functionality and Capabilities

Clef is a multimodal model that accepts a state and a schema of typed questions as input, then outputs decisions. It can process various data formats including text, JSON, images, and video. The model returns a probability for each allowed option for every question in a single pass, eliminating the need for free-form text generation or output parsing. This functionality allows AI agents to use the results to determine actions, such as routing support requests or deferring to human intervention.

Performance and Differentiation

Cloudflare also released Clef-Flash, a smaller 9B multimodal model optimized for latency-sensitive decisions. Clef-Flash achieved a median latency of 38.8 ms in Cloudflare's benchmarks, compared to 209.3 ms for the larger 27B Clef model. Cloudflare states that Clef differs from other decision models, such as Typesafe AI’s Jev System One, by including a vision encoder for classifying visual content and offering a larger 64k context window, compared to Jev's 32k.

Infrastructure and Industry Context

The models are hosted on Cloudflare’s infrastructure, utilizing GPUs at the edge to achieve low network latency and faster decision-making. This setup allows Clef to be integrated into critical paths for AI agents, potentially combining with Cloudflare's LLMs on Workers AI for taking action. While Cloudflare presents Clef as a new offering, some community discussions suggest that decision models are not a new paradigm, with Typesafe having previously marketed similar solutions.