AnswersPublished14 sources
Xiaomi-Lens-0: How Xiaomi’s 4B Model Grounds Answers in Image Pixels
Xiaomi Lens 0 is a 4B Qwen3.5 4B based model that can pair text with pixel masks for image regions, including during reasoning.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Xiaomi Lens 0 is a 4B Qwen3.5 4B based model that can pair text with pixel masks for image regions, including during reasoning.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Xiaomi-Lens-0, the 4-billion-parameter multimodal model open-sourced by Xiaomi’s SeerRay-Lab in October 2026, and how do its officia. Article summary: Xiaomi-Lens-0 is a compact 4-billion-parameter multimodal model built on Qwen3.5-4B that combines textual answers with pixel-level image-region masks, using a shared visual backbone and a lightweight mask decoder. [3] It. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
Xiaomi-Lens-0 is a 4-billion-parameter multimodal model designed to do more than describe an image: it can connect an answer to the pixels it refers to. Built on Qwen3.5-4B, it uses a shared visual backbone and a lightweight decoder to turn special mask tokens into image-region masks. The project says those grounded references can appear in a final answer or during reasoning. 1
How Xiaomi-Lens-0 produces image masks
The model learns from pixel-level supervision through a shared visual backbone. When it generates a <|mask|> token, a lightweight decoder converts that token into a mask over the image. This lets the model return text while also indicating which region supports its response. 1
The project describes this approach as grounded reasoning with visual latents. That points to region grounding being useful during reasoning, not only as a segmentation result at the end. The available project description does not spell out the full token protocol or exactly how intermediate masks affect later reasoning steps. 1
10
Supported tasks and LensBench
SeerRay-Lab lists referring and reasoning segmentation, multi-instance grounding, fine-grained region understanding, and visual perception among the project’s capabilities. 10 The model’s reported referring-expression results also cover eight RefCOCO-family splits.
1
The project lists LensBench as a segmentation benchmark, alongside its code, model, report, and project page. 10 News coverage also describes an accompanying demo, but the available source information does not establish its specific features or LensBench’s size and evaluation protocol.
2
What the reported results show
The project reports an average 5.6-point improvement in referring comprehension over Qwen3.5-4B across eight RefCOCO-family splits. It also says scores improved on all eight fine-grained perception benchmarks. These are results reported by the project, not independent reproductions. 1
A secondary report summary says CountQA improved by 6.8 points and that Xiaomi-Lens-0 had the highest segmentation cIoU among the systems compared. The summary does not provide enough detail to assess the exact baseline, individual scores, or full comparison set, so those claims should be read as reported results rather than a complete performance ranking. 2
Availability, licensing, and practical caveats
The model weights are listed on Hugging Face, and the project code is published on GitHub under SeerRay-Lab. 1
29 The GitHub repository identifies the project as Apache-2.0 licensed.
1 The Hugging Face page links the model to technical report 2610.23279 and shows a Transformers image-text-to-text pipeline example. That example alone does not establish whether the custom mask-decoding path works through a stock pipeline without additional project-specific setup.
29
Coverage attributes the release to Xiaomi, while the code and model repositories are published under the SeerRay-Lab name. The available sources do not independently establish the lab’s formal organizational relationship to Xiaomi. 2
1
29
Pixel-level grounding could be useful in applications such as identifying an object for a robot pick-and-place system, highlighting something a passenger asks about, or helping a phone user locate an item in a picture. These are potential use cases, not verified deployments. The available sources do not establish on-device speed, memory use, power consumption, or readiness for safety-critical settings.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Xiaomi-Lens-0: How Xiaomi’s 4B Model Grounds Answers in Image Pixels"?
Xiaomi Lens 0 is a 4B Qwen3.5 4B based model that can pair text with pixel masks for image regions, including during reasoning.