Xiaomi's Lens-0 4B MLLM Tops cIoU Across Eight RefCOCO Splits
TL;DR
- Xiaomi-Lens-0 is a 4B multimodal model on Qwen3.5 that generates mask tokens its own lightweight decoder turns into regions, with no separate segmentation encoder.
- Across the eight RefCOCO, RefCOCO+ and RefCOCOg splits, the paper reports a 5.6-point average gain over the Qwen3.5 backbone in referring comprehension.
- Weights are posted on Hugging Face under Apache 2.0 at a 5B-parameter BF16 footprint, with the technical report filed October 16, 2026.
Xiaomi's SeerRay Lab has released Xiaomi-Lens-0, a 4B multimodal model built on Qwen3.5 that the authors say posts the highest referring-segmentation cIoU across all eight RefCOCO, RefCOCO+ and RefCOCOg splits. The technical report, filed to arXiv on October 16, puts the gain over the Qwen3.5 backbone at 5.6 points on average in referring comprehension across those eight splits.
The design cuts out the usual separate segmentation head. "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," the paper reports. Each mask token emitted by the LM decodes to an image region, which the authors frame as letting the model "think with visual latents" during chain-of-thought.
On segmentation specifically, margins over the best competing score "range from 0.9 to 2.1 points on RefCOCO, 3.1 to 4.9 on RefCOCO+, and 2.8 to 2.9 on RefCOCOg," according to the paper. The retrieved abstract does not name which models those margins are measured against.
Weights are posted on Hugging Face under Apache 2.0 at a 5B-parameter total footprint in BF16, with evaluations spanning six tasks including reasoning segmentation and embodied referring.
Originally reported by paper
Read the original article →Original headline: Xiaomi SeerRay Ships 4B MLLM Topping Segmentation cIoU on All Eight RefCOCO Splits, Apache 2.0