Xiaomi-Lens-0 adds 5.6 points over Qwen3.5 on grounding splits

Xiaomi-Lens-0, a 4B multimodal model built on Qwen3.5, averages 5.6 points of gain over its backbone across the eight RefCOCO, RefCOCO+ and RefCOCOg comprehension splits, according to the Xiaomi-Lens-0 Technical Report.

The architectural claim is small and specific. The report says 'a lightweight decoder combines vision-encoder patch features with visual latents' so that 'mask supervision reaches the shared backbone without an additional segmentation encoder.' Each mask token in the language model carries a latent that the decoder resolves back to pixels; the authors frame this as a way to 'interleave reasoning with region-specific latents.'

The numbers follow the architecture. On the eight referring-expression segmentation splits the model averages 81.4 cIoU, ahead of a Qwen3-VL-Seg baseline at 78.6. On ReasonSeg Test it posts 68.0 gIoU; on RoboRefIt TestA it reports 87.0 gIoU and 88.4 cIoU. Fine-grained perception moves too: CountQA accuracy rises to 50.6 from a Qwen3.5 baseline of 43.8, a 6.8-point gain, and MME-RW-cn jumps 7.2 points.

The paper also introduces LensBench, a 1,200-unit, 4,800-instruction benchmark crossing single and multiple targets with explicit and implicit wording across general, aerial, driving and panoramic images. On the panoramic multi-target implicit slice, Xiaomi-Lens-0's [email protected] drops from 72.1 to 48.7, which the authors flag as the model's largest decline. The retrieved report publishes no training compute, no inference-latency numbers, and no weights-license detail.

It lands on a crowded week for open releases: our open-source tracker has logged 357 stories over the last 90 days.