Xiaomi opens Lens-0, a 4B model that segments what you describe
Built on Qwen3.5, it emits masks from the LM itself — no segmentation head — tops cIoU on all eight RefCOCO splits, and ships under Apache 2.0.
Copy markdown
A 4B model that points at anything you name
Lens-0 does referring-expression segmentation, multi-instance grounding, reasoning segmentation, and embodied referring from plain text: describe a region, get the mask back. It's built on Qwen3.5-4B and runs on a single CUDA GPU.
Masks come out of the LM, not a bolted-on head
There's no separate segmentation encoder — the language model emits mask tokens that a lightweight decoder turns into image regions, so a mask can surface mid-reasoning. The authors call it "thinking with visual latents."
The scores: top cIoU on all eight RefCOCO splits
Lens-0 reports the best referring-segmentation cIoU across every RefCOCO, RefCOCO+, and RefCOCOg split, and gains +5.6 points on average over its Qwen3.5 backbone in referring comprehension. Margins over the next-best run reach 4.9 points on RefCOCO+.
Build this weekend: a self-hosted 'point at it' API
A 4B grounding model you own means UI-agent targeting, bulk image auto-labeling, or a describe-to-crop tool with no per-call vision bill. For serving, vLLM splits into two ports — text generation on 8000, mask decoding on 8001.
Apache 2.0, with a community int8 build for phones
Weights are Apache 2.0 on Hugging Face (the Qwen3.5-4B base is too), so commercial use is clean. An MNN int8 conversion is already up for on-device inference; the BF16 path needs FlashAttention-2. Fair warning: the card says it can miss small or occluded objects.