How xiaomimimo/mimo works
MiMo-7B is a reasoning-model project documented from pretraining through supervised fine-tuning and reinforcement learning. Its README describes data preparation, training objectives, rule-based rewards, and rollout infrastructure; the only sampled runtime source registers a MiMo causal-language model implementation with vLLM. Training-stage links below reflect README statements, not inspected training code, so their nodes have no implementation path. The repository does not show the full training or serving workflow.
Main entry points: register_mimo_in_vllm.py MiMo model adapter
Components 16
Pretraining
- Reasoning data
- Base pretraining
- MTP objective
- Base model
Posttraining
- Supervised tuning
- RL problems
- Rule verifiers
- Code rewards
- Easy resampling
- Rollout engine
- RL model
- SFT model
Inference
- MiMo model adapter
registry/register_mimo_in_vllm.py - vLLM runtime
Other components
- Model developers
- Model users
Connections 15
- Model developers → Reasoning data: curate data
README.md - Reasoning data → Base pretraining: trains on
README.md - Base pretraining → Base model: produces
README.md - MTP objective → Base pretraining: adds objective
README.md - Base model → Supervised tuning: starts tuning
README.md - Supervised tuning → SFT model: produces
README.md - Base model → Rollout engine: starts RL
README.md - SFT model → Rollout engine: starts RL
README.md - RL problems → Rule verifiers: verified by
README.md - Rule verifiers → Rollout engine: scores answers
README.md - Code rewards → Rollout engine: shapes rewards
README.md - Easy resampling → Rollout engine: improves sampling
README.md - Rollout engine → RL model: trains
README.md - MiMo model adapter → vLLM runtime: registers model
registry/register_mimo_in_vllm.py - Model users → vLLM runtime: request inference