Xiaomi Opens Xiaomi-Lens-0, a 4B MLLM That Decodes Pixel Masks From Qwen3.5's Own Visual Features

Xiaomi's technical report introduces Xiaomi-Lens-0, a 4B multimodal LLM built on Qwen3.5 that decodes region masks directly from the shared vision encoder's patch-level features — no separate segmentation head. It averages +5.6 points over the Qwen3.5 baseline on REC splits, tops cIoU on all 8 RefCOCO/+/g RES splits (81.4 avg), hits 68.0 gIoU on ReasonSeg, and gains +6.8 points on CountQA. The authors also release LensBench, a new human-reviewed grounding benchmark.