Xiaomi-Lens-0
Xiaomi released Xiaomi-Lens-0, a 4B multimodal model built on Qwen3.5 that predicts segmentation masks from its own visual representations and gains 5.6 points on RefCOCO over Qwen3.5.
A lightweight decoder combines vision-encoder patch features with visual latents, which are the language model's mask-token states. Mask supervision therefore reaches the shared backbone without a separate segmentation encoder. Each latent decodes to a mask, so the model can interleave reasoning with region-specific latents, and grounded reasoning traces supervise the regions it visits during chain-of-thought.
- Date
- Friday 16 October 2026
- Lab
- Xiaomi
- Kind
- paper
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Average gain over Qwen3.5 in referring comprehension across eight RefCOCO, RefCOCO+ and RefCOCOg splits | 5.6 points also reports the highest segmentation cIoU among compared models and beats Qwen3.5 on all perception benchmarks | company |
The abstract mentions a homepage, but weights and code availability were not confirmed.
Sources
This record was checked against its sources on 19 October 2026. How we check