Cloudflare releases open-source Clef decision models for fast agentic routing and bot classification

Summary

Cloudflare launched Clef and Clef-flash, two open-source non-autoregressive decision models hosted on Workers AI and available under an Apache 2.0 license on Hugging Face. Built on frozen Qwen backbones with vision capabilities and a 64k context window, the models output structured probabilities rather than free-form text. Cloudflare is using the architecture internally for threat intelligence and bot detection while rolling out reinforcement learning fine-tuning services for enterprise agent workflows.

Insight

By replacing token-by-token autoregressive generation with parallel schema scoring over base model representations, decision models dramatically compress inference latency and eliminate non-deterministic parsing issues in automated routing and security triage pipelines.

Implication

Autonomous web agents and edge security filters can execute programmatic decisions—such as tool selection, request routing, or bot classification—at edge latency without routing traffic through heavyweight generative LLMs.

Why it matters

As web scraping, bot defense, and agentic interactions accelerate, deploying sub-100ms structured classifier models directly to edge infrastructure shifts how platforms detect automated traffic and how autonomous agents navigate the web.

Evidence

• Primary source
• Evidence ID: cloudflare-blog--962aa10ad52cf1e3