MiMo VL 7B SFT
MiMo VL 7B SFT is the result of four stages: projector warmup, vision-language alignment, multimodal pretraining and long-context supervised tuning.
Research release
Reasoning data enters before the final tuning
Xiaomi includes synthetic reasoning data during later pretraining stages, rather than reserving it only for a final fine-tune. The SFT weights allow study of that training stage separately from RL.MiMo-VL-7B-SFT official model card
Use it as a visual adaptation baseline
Practical guidance. A supervised checkpoint can help compare different visual reward schemes. Check image grounding and factual extraction before and after adaptation. Better reasoning prose is not sufficient if the model points to the wrong object.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | MiMo-VL-7B-SFTMiMo-VL-7B-SFT-GGUF |
| Inputs | Text, images and video in the published vision-language model family.MiMo-VL-7B-SFT official model card |
| Licence | MIT.MiMo-VL-7B-SFT official model card |
| Stage | This is the supervised checkpoint before the separate RL release.MiMo-VL-7B-SFT official model card |
Related models
- MiMo VL 7B RL
Xiaomi · MiMo VL
MiMo VL 7B RL continues the supervised vision-language checkpoint with mixed on-policy reinforcement learning.
Launch date not confirmedDownloadable - MiMo VL 7B SFT 2508
Xiaomi · MiMo VL
An updated MiMo VL 7B instruction model with a selectable reasoning mode.
Launch date not confirmedResearch releaseDownloadable