Xiaomi has open-sourced Xiaomi-Lens-0, a 4-billion-parameter multimodal model that can mark the exact image region it is talking about while it answers. Weights are on Hugging Face under SeerRay-Lab, code is on GitHub, and a technical report is on arXiv. The release is Apache 2.0.
SeerRay-Lab is the Visual Understanding Group inside Xiaomi's large language model applications team. The model card and report credit Xiaomi Corporation for the Xiaomi-Lens-0 modifications, so this is a Xiaomi research release hosted under the lab's Hugging Face organization, not a third-party fork of a Xiaomi name.
The model is built on Qwen3.5-4B and adds a lightweight pixel-mask decoder. When it emits a special mask token, that token's visual latent is decoded into a segmentation mask over the image. Because each latent maps to a region, the model can interleave ordinary text with region pointers during chain-of-thought, not only at the final answer.
Supported tasks include referring expression comprehension and segmentation, multi-instance grounding, reasoning segmentation, embodied referring such as hand-me-the-object prompts, and fine-grained visual perception. The team also released LensBench, a segmentation benchmark, plus a demo.
On the eight RefCOCO, RefCOCO+ and RefCOCOg referring splits, Xiaomi-Lens-0 gains 5.6 points on average over the Qwen3.5 base in referring comprehension, according to the technical report. The report also cites a CountQA gain of 6.8 points and says the model leads compared systems on segmentation cIoU among the models it lists. Those scores are the authors' evaluations.
At about 4B parameters, the model is small enough for edge-oriented experiments in robotics pick-and-place, in-car assistants and phone visual helpers that need to show where an answer applies. The release ships weights, config, tokenizer and processor; loading and mask decoding use the accompanying Xiaomi-Lens package rather than a stock Transformers class alone.
SEE ALSO: Xiaomi Open-Sources MiMo-V2.6 Pro and Flash · SenseTime SenseNova U1.5 Brings Native Unified Vision