Xiaomi-Lens-0: Fine-Grained MLLM with Mask-Token Reasoning
October 15, 2026
Xiaomi-Lens-0 is a 4B MLLM based on Qwen3.5 that predicts masks directly from its visual representations. By combining vision-encoder patch features with mask-token states, the model can interleave reasoning with region-specific visual latents.
HOW THIS AFFECTS YOU
●
builderYou can build multimodal applications that require precise, grounded reasoning about specific image regions.
●
researcherThis architecture shows how to integrate segmentation capabilities directly into the language backbone.