Cloudflare launches Clef open-weight multimodal decision models

Cloudflare has launched Clef and Clef-flash, two open-weight multimodal decision models designed for tasks such as structured yes/no, multiple-choice, and ranking. The models utilize Qwen as their backbone, specifically using versions of Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash.

Unlike some competitors like TypeSafe’s Jev, which focuses on text, Clef is natively multimodal, capable of processing images and video. The models also feature an expanded context window of up to 64k, allowing for more comprehensive system inputs. Cloudflare reports that Clef outperforms Jev on several benchmarks, including the Jev Decision Index, though the company’s self-reported scores have yet to be independently reproduced.

Cloudflare also introduced a new reinforcement learning product that enables customers to perform customized fine-tuning of the Clef models for specific business scenarios. The models are available for local use via Hugging Face and are compatible with the Jev-API.