Cloudflare has released two new open-source decision models, Clef and Clef-flash, available on Workers AI. Unlike traditional Large Language Models (LLMs) that generate text token by token, these decision models are designed to produce fast, deterministic, and structured outputs to help AI agents make programmatic decisions in various workflows.
Cloudflare Releases Clef: Open-Source Decision Models for Fast, Structured Agentic Workflows
The decision models work by classifying inputs—such as a customer support message—and returning typed answers with associated probabilities. This allows developers to automate tasks like ticket routing or escalation without requiring a human in the loop for every step. For example, Cloudflare’s Threat Intelligence team uses Clef to classify website domains; the model can identify a domain's category (e.g., ecommerce or phishing) in 2.2 seconds, significantly faster than a general LLM.
The Clef family includes two versions: the standard Clef model for high-precision tasks and Clef-flash for latency-critical applications. Both models feature a 64k context window and a vision encoder, enabling them to classify visual content as well as text. Additionally, the models are fully Jev-API compatible, allowing for easy integration into existing workflows.
Cloudflare is also debuting a new reinforcement learning (RL) product, which will allow customers to fine-tune Clef models for specific use cases through a hands-on service or a self-serve platform. Currently, the models are being released under the Apache 2.0 license on Hugging Face for local experimentation.
Sources
- Clef: our open-source decision models （Hacker News Frontpage, 2026-10-01）