Olive Song, Senior Researcher at MiniMax, presents the MiniMax M2 model, a language model designed specifically for coding and agentic tasks. With only 10 billion active parameters, it positions itself as an extremely capable and "cost-efficient" alternative to giant models, targeting developers and enterprises in particular.
The model's strength rests on several key innovations in its training: 1. Realistic coding experience: MiniMax uses its own expert developers as Reward Models for Reinforcement Learning, aligning the model's behavior with the real expectations of engineers (code quality, reliability). 2. Interleaved Thinking: To handle long-horizon tasks, M2 does not use a simple linear chain of thought. It dynamically alternates between "Thinking" and "Acting" (using a tool). If a tool fails or returns an unexpected result (environment noise), the model reassesses the situation and attempts a different approach, mirroring human behavior when facing uncertainty. 3. Robust generalization: To prevent the agent from performing well only within a specific setup, MiniMax injects constant perturbations into the training data (changes to prompt format, tool response format). This makes the agent capable of adapting to different "scaffolds" (execution environments) without losing effectiveness. 4. Multi-agent scalability: The model's small size allows several instances to run in parallel (e.g., a research agent, a writing agent, a front-end agent) to solve complex tasks collaboratively at lower cost.
Olive Song concludes by presenting the future roadmap (M2.5, M3), including improved context and memory management, and native multimodal integration (audio/video).