Cloudflare Launches Clef, Open-Source Multimodal Decision Models Based on Qwen
Cloudflare has introduced Clef and Clef-flash, open-source multimodal decision models built on Alibaba's Qwen architecture. The models feature native image understanding, an expanded 64k context window, and launch alongside a new reinforcement learning product for custom fine-tuning. By expanding AI decision-making tools to process visual inputs alongside text within a large 64k context, Cloudflare enables developers to build more context-aware classification and routing systems. Providing these models open-source with tailored fine-tuning tools simplifies deploying specialized AI agents in edge and enterprise workflows. Clef is based on Qwen3.8-27B, whereas Clef-flash is built on Qwen3.5-9B, both incorporating native vision encoders to evaluate visual content directly. Both models maintain full compatibility with Jev-API and outperform baseline benchmarks such as the Jev Decision Index 0.2.1.
## BACKGROUND
AI decision models focus on parsing structured complex inputs to select optimal actions or classifications rather than generating free-form conversational text. Open-weight base LLMs like Alibaba's Qwen serve as flexible foundations for specialized fine-tuning across multimodal tasks.