The underpinnings of Cloudflare's recently released clef decision models rely on the Qwen architecture from Alibaba, with weights available on Hugging Face. The most efficient variant, clef-flash, uses a Qwen3.5-9B base to operate approximately 13 times faster than Jev, while the standard Clef model is built on Qwen3.8-27B. Both models are multi-modal and maintain compatibility with the Jev API.

To support the deployment of these first in-house models, Cloudflare integrated Cog, a toolset from Replicate, into its Workers AI platform. This integration enables developers to fine-tune and redeploy the clef models within the provider's environment to optimize for latency and quality.