QwenVL3.5-4B

QwenVL3.5-4B is a 4-billion-parameter-class vision-language model from the Qwen3.5-VL family developed by Alibaba. Under this name it appears primarily in the technical documentation of the Chinese company PsiBot, which used it as the upper-layer (planning) backbone of its Psi-R2.5 embodied AI foundation model, unveiled in September 2026.

In that role the model processes vision-language instructions and decomposes them into subtasks, drawing on metadata context — memory and soft prompts — and on value signals from reinforcement learning. Its output feeds the lower layer built on Wan2.2-TI2V-5B, which turns the plan into robot action trajectories. PsiBot states that both layers leverage its own proprietary pre-training data.

A note on naming: the Qwen3.5-VL family does exist and is published by Alibaba in 2026 in several sizes (including 2B, 9B and 122B-A10B checkpoints, with the 9B version serving as the basis for Qwen-Image-2.1). Alibaba has not, however, published an official model card under the exact name "QwenVL3.5-4B" — that spelling comes from PsiBot's materials, which remain the only direct source for the 4B variant used in Psi-R2.5. The exact parameter count, context window and licence of this particular checkpoint have not been disclosed publicly.