Clef: our open-source decision models (blog.cloudflare.com)

Cloudflare has announced the release of Clef and Clef-flash, open-source decision models that enhance the efficiency and accuracy of decision-making in AI workflows. Unlike traditional classifiers and Large Language Models (LLMs), Clef produces quick, deterministic outputs without the need for constant retraining. It excels at generating structured, probabilistic outputs — for example, classifying a website domain with high accuracy in a fraction of the time compared to existing models like Jev System One. Clef's architecture allows it to operate with a context window of 64k, double that of its competitors, making it adaptable for a variety of input types, including images. This release is significant for the AI/ML community as it introduces new capabilities for programmatic decision-making, potentially reducing human involvement in routine tasks. Clef models leverage Cloudflare's infrastructure for low-latency performance, enabling rapid decision-making in applications ranging from threat intelligence to customer service. The addition of a reinforcement learning (RL) fine-tuning service will further enhance these models, allowing tailored adjustments for specific enterprise needs. Open-sourcing Clef under an Apache 2.0 license on Hugging Face encourages widespread experimentation and collaboration in the development of efficient AI decision systems.