MiMo VL 7B RL
MiMo VL 7B RL continues the supervised vision-language checkpoint with mixed on-policy reinforcement learning. Its rewards cover perception, grounding, reasoning and preferences.
Mixed objectives can interfere
The authors explicitly note that interference between training domains remains a challenge. Their method tries to improve several visual and reasoning abilities together.MiMo-VL-7B-RL official model card
Evaluate localization as well as explanation
Practical guidance. Ask for the object or region that supports an answer, then check it against the image. This can reveal a plausible explanation attached to the wrong evidence. Use video tasks that require timing, not just a summary of one frame.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MiMo-VL-7B-RLMiMo-VL-7B-RL-GGUF |
| Licence | MIT.MiMo-VL-7B-RL official model card |
| Training distinction | The reward mix spans text, image and video tasks rather than one narrow visual skill.MiMo-VL-7B-RL official model card |
Related models
- MiMo VL 7B RL 2508
Xiaomi · MiMo VL
An updated MiMo VL 7B reasoning model for image and video tasks with an option to disable reasoning.
Launch date not confirmedDownloadable - MiMo VL 7B SFT
Xiaomi · MiMo VL
A 7B vision-language model trained through alignment, multimodal pretraining and supervised instruction tuning.
Launch date not confirmedResearch releaseDownloadable