Xiaomi Open-Sources 4B Research Model That Points at Image Regions

Xiaomi-Lens-0 can answer a question about an image and, in the same breath, outline the exact area of the picture it is talking about. The release, published on 21 October 2026, comes from SeerRay-Lab, the Visual Understanding Group inside Xiaomi's large language model applications team. The model card and report credit Xiaomi Corporation, so this is a Xiaomi research release, not a third-party copy.

The model is built on Qwen3.5-4B and adds a lightweight pixel-mask decoder. When it emits a special mask token, the decoder turns that token's visual representation into a segmentation mask over the image. Because each representation maps to one region, the model can interleave ordinary text with region pointers during its step-by-step reasoning, not only in the final answer. The team also released LensBench, a segmentation benchmark, and a demo.

The technical report credits the model with an average gain of 5.6 points over the Qwen3.5 base on referring comprehension, the task of finding the object a sentence describes, across eight splits of the RefCOCO, RefCOCO+ and RefCOCOg tests. It also cites a 6.8-point gain on CountQA and says the model leads the systems it lists on segmentation cIoU, a measure of how closely predicted masks match the true ones. These are the authors' own evaluations, and the release mentions no independent test.

For developers and researchers, the change is access: the weights, code and report are public under the Apache 2.0 license, which allows reuse. The release names robotics pick-and-place, in-car assistants and phone visual helpers as edge-oriented experiments, since a 4-billion-parameter model is small enough for such devices. None of these uses is described as a product, and no date is given. Loading the model requires the accompanying Xiaomi-Lens package, not a standard Transformers class alone.