Xiaomi releases 'Xiaomi-Lens-0', a 4B multimodal model that points to image regions at pixel level mid-answer, plus the 'LensBench' benchmark (report)

Verified (fact)

According to Pandaily, Xiaomi's SeerRay-Lab open-sourced 'Xiaomi-Lens-0', a 4B multimodal model built on Qwen3.5-4B. While generating an answer, the model emits mask tokens that mark, at the pixel level, which region of the image it is talking about. The team also released 'LensBench', a benchmark for evaluating this capability.

Why it matters (interpretation)

A small model answers in text while precisely segmenting and pointing to image regions at the same time. That could help in tasks where it matters what the model is looking at, such as on-device visual assistants and robot perception.

Sources

• Press Xiaomi Open-Sources Xiaomi-Lens-0, a 4B Model That Points to Image Regions Pixel by Pixel Pandaily

More about Xiaomi

• 2026-10-22 P2 Model API Language modelsAgents XiaomiTencent
  Xiaomi's MiMo-V2.6-Pro comes to Tencent's WorkBuddy agent and TokenHub with 1M-token context (report)

  Pandaily reported that Xiaomi's MiMo-V2.6-Pro is available inside Tencent's WorkBuddy agent and on Tencent Cloud TokenHub under the name...

• 2026-10-16 P2 Open source Open weights Speech & audioMultimodal Xiaomi
  Xiaomi open-sources MiDashengLM-Spatial under Apache 2.0, an audio model that understands what a sound is and its direction, distance, and movement (report)

  According to Pandaily, Xiaomi has open-sourced 'MiDashengLM-Spatial,' a spatial audio understanding model, under the Apache 2.0 license....