Cloudflare has released Clef and Clef-flash, two open-source decision models designed to help AI agents make fast, consistent classifications without constant retraining. The models are available on Workers AI and open-sourced on Hugging Face under an Apache 2.0 license.
Unlike Large Language Models, which are non-deterministic and open-ended, decision models produce bounded, structured outputs with probabilities—enabling agents to make autonomous decisions and take actions without human intervention. For example, a customer support ticket can be automatically classified as urgent or routine and routed to the appropriate team based on typed probability scores.
Cloudflare tested Clef on its Threat Intelligence team to classify website domains. The model took 2.2 seconds to fetch, render, and classify a domain—for example, identifying a site as 95% likely a fashion website and 85% ecommerce—while the fastest general LLM tested (gpt-oss-120b) took 4.7 seconds and returned only two classifications. According to Cloudflare, this 2x latency improvement enables faster identification of malicious or legitimate domains.
Clef differentiates itself from competing decision models in three ways. First, it includes a vision encoder, allowing it to classify images as well as text. Second, it has a 64k context window, compared to Jev’s 32k, enabling models to process more input state. Third, according to benchmarks evaluated against the Jev Decision Index, Clef scores competitively across various quality metrics, with Clef-flash performing exceptionally well despite being significantly faster than other models.
On latency benchmarks across 43 evaluations, Clef models beat other decision models except Laya, which trades off quality for speed. Clef’s median latency was 209.3 milliseconds, while Clef-flash achieved 38.8 milliseconds. Being hosted on Cloudflare’s edge infrastructure with GPUs provides additional network latency benefits.
The models produce strictly typed outputs and are fully API-compatible with Jev, allowing easy migration. Cloudflare guarantees it does not read, store, or train on requests or responses unless customers opt into the company’s new reinforcement learning fine-tuning product, which allows customization of Clef for specific use cases. The models are enterprise-ready and available through developer documentation or the open-source Hugging Face repository.
Key facts
- Cloudflare released Clef decision models on Workers AI, open-sourced on Hugging Face under Apache 2.0 license
- Decision models produce deterministic, typed classifications with probabilities for autonomous agent decision-making
- Clef achieved 2.2 seconds latency on domain classification versus 4.7 seconds for gpt-oss-120b LLM
- Clef includes a vision encoder for image classification and has a 64k context window, larger than competing models
- Median latency of 38.8ms for Clef-flash and 209.3ms for standard Clef across 43 benchmarks
- Cloudflare offers reinforcement learning fine-tuning to customize Clef for specific workflows