Cloudflare Ships Open-Source Clef Models for Automated Support and Ops Triage

Cloudflare has introduced two open-source decision models called Clef and Clef-flash on its Workers AI platform, complemented by a reinforcement learning (RL) fine-tuning product. Unlike large language models, decision models generate bounded, typed, probability-scored outputs for specific questions, allowing workflows to act directly on the output without parsing free text.

An example provided by Cloudflare is to input a support message and receive structured, probability-weighted answers to determine urgency, responsible team, and impact severity for ticket routing or human escalation.

Clef currently holds the top spot on the Jev Decision Index benchmark and is API-compatible with Typesafe AI's Jev models, enabling existing integrations to replace Jev models with Clef without requiring modifications. Both models are released on Hugging Face under the Apache 2.0 license, allowing local usage, and are also available on Workers AI for immediate API-based access.

Performance-wise, Clef classifies a website domain faster, taking 2.2 seconds and providing more category labels compared to 4.7 seconds for the general-purpose LLM gpt-oss-120b, which yields only two classifications. Clef also incorporates a vision encoder for image classification, a feature absent in Jev. Furthermore, Clef offers a 64k context window, surpassing Jev's 32k limit.

Alongside the models, Cloudflare is unveiling an RL fine-tuning service. Initially, this service involves personalized sessions with Cloudflare's forward-deployed engineer (FDE) team, while a self-serve version is scheduled for later availability, enabling customers to capture their own request data via Cloudflare's AI Gateway, train on Workers AI, and redeploy a fine-tuned Clef model on Cloudflare's infrastructure.

Importantly, Cloudflare assures that it does not read, store, or train on customer requests unless customers explicitly opt into the fine-tuning service.