Clef is based on the Qwen3.8-27B model and Clef-flash is based on the Qwen3.5-9B model, with both keeping their respective base vision encoders