Multimodal LLMs can describe whole images but often fail when asked to pinpoint a specific region. Xiaomi-Lens-0, a 4B model built on Qwen3.5, tackles this by making the language model itself produce masks. A lightweight decoder combines vision-encoder patch features with visual latents—the model's mask-token states—so mask supervision reaches the shared backbone without adding a separate segmentation encoder.
Because each latent decodes to a mask, the model can interleave reasoning with region-specific latents, and grounded reasoning traces supervise the regions it visits during chain-of-thought. This lets the model "think with visual latents" rather than treating segmentation as a post-hoc step.
On eight RefCOCO, RefCOCO+, and RefCOCOg splits, Xiaomi-Lens-0 gains 5.6 points on average over Qwen3.5 in referring comprehension and achieves the highest segmentation cIoU among compared models. It also surpasses Qwen3.5 on all perception benchmarks tested. The authors report these results in a technical report on arXiv; no independent evaluation is included.