Xiaomi Open Sources a 4B Model That Points at Pixels
Xiaomi has released Xiaomi Lens 0, a four billion parameter model that can mark the exact region of an image it is talking about while it answers, with the weights, code and technical report published openly. The work comes from SeerRay Lab, the visual understanding group inside Xiaomi's large language model applications team, and it is licensed under Apache 2.0.
The model sits on Qwen3.5 4B and adds a lightweight decoder for pixel masks. When it emits a special mask token, that token's visual latent is turned into a segmentation mask over the image, so the model can point at a region in the middle of a chain of thought rather than only in a final answer. According to Pandaily, the weights sit on Hugging Face under the lab's organisation, the code is on GitHub and the technical report is on arXiv, which puts the whole release in public view rather than behind a request form.
And the range of tasks is wider than a single pointing trick. The team lists referring expression comprehension and segmentation, multi instance grounding, reasoning segmentation, embodied referring for prompts such as handing over an object, and fine grained visual perception. It also published LensBench, a segmentation benchmark, along with a demonstration.
On the eight referring splits of RefCOCO, RefCOCO+ and RefCOCOg, the model gains 5.6 points on average over the Qwen3.5 base in referring comprehension, according to the report, which measures a task where a sentence has to be matched to the part of a picture it describes. The same report cites a gain of 6.8 points on CountQA and says the model leads the systems it lists on segmentation cIou. Those scores are the authors' own evaluations and have not been reproduced independently.
At about four billion parameters the model is small enough to run in edge experiments. The team names robotics pick and place, in car assistants and phone visual helpers as settings where a system needs to show where an answer applies rather than describe it in words alone.
But loading it takes more than a stock class from the Transformers library. The release ships weights, configuration, tokenizer and processor, and both loading and mask decoding depend on an accompanying Xiaomi Lens package, so a team that expects a drop in replacement will have to add the repository to its stack.
The choice of Qwen3.5 as a base is worth noting. Xiaomi built its pointing layer on another company's open model rather than training a vision system from scratch, which keeps the compute cost low and puts the contribution in the decoder and the data used to train it.
So the practical question is how well the masks hold up outside the benchmark. Segmentation scores are measured against curated sets, and a phone assistant or a robot arm has to cope with cluttered scenes, poor light and objects that shift between the question and the answer.
So the open questions sit on the record rather than in a footnote. Xiaomi has published no latency figure for on device use, and the report gives no detail on the size of the training set behind the pointing data. The licence is permissive, which means other teams can test those gaps themselves.
Xiaomi has open sourced Xiaomi Lens 0, a four billion parameter model that marks image regions as it answers.
It is built on Qwen3.5 4B with a pixel mask decoder, and released under Apache 2.0 with weights, code and a report.
The team reports a 5.6 point average gain on eight referring splits against the base model.
Scores come from the authors' evaluations and have not been reproduced independently.
The model needs its own package to load and decode masks, so it is not a drop in replacement.
Source: Pandaily
