Xiaomi MiMo-V2.5 310B Architecture
Back to Xiaomi MiMo-V2.5 310B in the gallery
Xiaomi
Omnimodal sparse MoE model that extends the MiMo-V2-Flash backbone with vision and audio encoders.
Architecture and specifications
- Scale
- 310B total, 15B active (4.8% active)
- Context (tokens)
- 1,048,576
- License
- MIT License
- Date
- 2026-04-22
- Decoder type
- Sparse omnimodal MoE
- Attention
- GQA with 8 sliding-window and 4 global KV heads. Keys have 192 dimensions and values have 128.
- Layer mix
- 39 sliding-window + 9 global
- KV cache / token (bf16) info
- 217.5 KiB · High
- Key detail
- Adds native image, video, and audio encoders to the MiMo-V2-Flash-style sparse MoE backbone with MTP support.
Training and benchmarks
- Pretraining data
- ~48T tokens Reported total for text and multimodal pretraining. Model card
- Post-training methods
- SFT, RL, Distillation SFT, agentic RL, and multi-teacher on-policy distillation (MOPD). Model card