ResearchResearch paperMultimodal Models · Reasoning & Planning · Large Language Models2 sources · Oct 16, 2026

Xiaomi-Lens-0 Technical Report

To close this gap, we present Xiaomi-Lens-0, a 4B MLLM built on Qwen3.5 that predicts masks from its own visual representations.

Multimodal large language models (MLLMs) answer questions about whole images well, yet they still struggle with fine-grained region understanding.
Specifically, a lightweight decoder combines vision-encoder patch features with visual latents, the language model's mask-token states, so mask supervision reaches the shared backbone without an additional segmentation encoder.
Because each latent decodes to a mask, the model can interleave reasoning with region-specific latents and thus think with visual latents, while grounded reasoning traces supervise the regions it visits during chain-of-thought.
As a result, across all eight RefCOCO, RefCOCO+, and RefCOCOg splits, Xiaomi-Lens-0 gains 5.6 points on average over Qwen3.5 in referring comprehension and achieves the highest segmentation cIoU among compared models.

[1]Xiaomi-Lens-0 Technical Report
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 10:50 AM
To close this gap, we present Xiaomi-Lens-0, a 4B MLLM built on Qwen3.5 that predicts masks from its own visual representations.
Multimodal large language models (MLLMs) answer questions about whole images well, yet they still struggle with fine-grained region understanding.
[2]Xiaomi-Lens-0 Technical Report
Hugging Face Daily Papers · Oct 16, 12:00 AM · same content

Extractive summary: sentences quoted from the sources.

Before this