Xiaomi open-sources Xiaomi-Lens-0, a 4B model that points to image regions pixel by pixel

Xiaomi has released Xiaomi-Lens-0, a 4-billion-parameter multimodal model that marks the exact image region it refers to while answering. Weights are on Hugging Face under Apache 2.0, code is on GitHub and a technical report is on arXiv. Built on Qwen3.5-4B with a lightweight pixel-mask decoder, it gains 5.6 points on average across RefCOCO splits, per the authors.

• 4B-parameter model built on Qwen3.5-4B with a pixel-mask decoder
• Gains 5.6 points on eight RefCOCO splits and 6.8 on CountQA
• Weights, code and technical report released under Apache 2.0
• Ships LensBench segmentation benchmark and a demo