Back to Gap Map Model components / Multimodal models

MiMo-VL

Xiaomi
open weights / Overall score: 2.7

MiMo-VL is Xiaomi's 7B vision-language model line, joining a native-resolution vision encoder to the MiMo-7B language model. It answers questions about images, documents and video and reasons step by step, with a switch that turns the long reasoning off. The 2508 release, the current one, ships as SFT and RL checkpoints and raised the line's image and video benchmark scores.

The repository's LICENSE is Apache-2.0 while the model cards say MIT; both are permissive, and the cards govern the weights.

Openness

3 high confidence
3.0
weights
open (MiMo-VL-7B-RL-2508 and SFT-2508 on the Hub, ungated)
data
described (pretraining and RL data described in the technical report, not released)
code
partial (inference demo and notebooks only)
license
MIT (weights)+Apache-2.0 (repository code)

The 2508 checkpoints are MIT and the repository code is Apache-2.0. Xiaomi describes its training stages and data in the report but publishes neither the data nor the training code, only inference examples.

Adoption

2 medium confidence
2.0

Hugging Face downloads over the trailing 30 days for Xiaomi's six MiMo-VL checkpoints, including its own GGUF builds, most of them for the 2508 RL model.

Capability

3 medium confidence
3.0

MiMo-VL reads images, documents and video, level with MiniCPM-V. Xiaomi reports it locating elements on a screen but gives no results for completing tasks in a live interface.

Verified 2026-11-12