Xiaomi Open-Sources Xiaomi-Lens-0, a 4-Billion-Parameter Multimodal Grounding Model
Xiaomi has open-sourced Xiaomi-Lens-0, a 4-billion-parameter multimodal model that can mark the exact image region it is referring to while answering, with weights on Hugging Face under the SeerRay-Lab organization, code on GitHub and a technical report on arXiv, released under Apache 2.0. Built on Qwen3.5-4B with a lightweight pixel-mask decoder, the model interleaves text with region pointers during chain-of-thought and supports referring-expression comprehension and segmentation, multi-instance grounding, reasoning segmentation and embodied referring prompts. At roughly 4B parameters it is small enough for edge experiments in robotics pick-and-place, in-car assistants and phone visual helpers, and it ships with a LensBench segmentation benchmark and demo.
Source: Pandaily / Hugging Face (SeerRay-Lab), Oct 21, 2026 | 2026-10-21
- Ant's Bai Ling and Kunlun Wanwei Open-Source New Models at Sustainability Leaders Conference
- Chinese Scientists Reportedly Use a Homegrown LLM to Crack a World-Class Math Problem
- U.S. Moves to Bar DeepSeek From Federal Use; Chinese Lab Doubles Down on Open Source
- SenseTime Open-Sources SenseNova-RoboRSI Embodied-AI Harness and 27B Robot-Brain Model
- Xiaomi Open-Sources Xiaomi-Lens-0, a 4-Billion-Parameter Multimodal Grounding Model
- China Issues Top-Level Plan to Deepen Tech-Industry Fusion, Launching Full 'AI+' Action
- Eight Ministries' 15th Five-Year Plan Backs General and Industry LLMs, Agent Decision Models
- China Proposes Export Controls on 40 AI, Biomedicine and PV Technologies; Open Source Exempt
- 2026 'Gold-Trade Global' Matchmaking for Middle East and ASEAN Opens at Yiwu Fair
- Zhejiang Lands Six Projects in First National Landmark Application Scenarios, Including Yiwu Digital-Trade Hub
