← back
Minimax M2: Building the #1 Open Model – Olive Song, MiniMax
Takeaway
MiniMax M2 shows that a small, cheap open model trained on perturbed agent scaffolds with expert-developer reward signals can rival larger closed models for coding and tool-use.
Summary
- Olive Song presents MiniMax M2, an open-weight model with only 10B active parameters targeted at coding and agentic workflows; hit #1 in downloads its first week and top-3 token usage on OpenRouter.
- Training scaled verifiable coding environments plus 'expert developer' reward models — in-house senior devs author problems, bug-fixes and refactors and grade outputs to shape behaviors developers trust.
- Uses 'interleaved thinking' so the model reasons between tool calls (tens to ~100 turns per user request), letting it adapt to noisy tool errors and stay stable on long-horizon tasks like a stock-perturbation demo.
- Generalization across agent scaffolds achieved via a perturbation pipeline that varies tool schemas, system prompts, chat templates and tool responses during training, rather than just scaling tool count.
- Small size enables multi-agent scalability — parallel M2 copies do research, writing and front-end work inside MiniMax's own agent app.
foundation-modelsopen-weightsagents
Original description
Introducing Minimax's latest AI model and its applications in code generation. Speaker: Olive Song | Senior Researcher, MiniMax https://x.com/olive_jy_song