Descriptions:
Olive Song, an RL and evaluation researcher at MiniMax, introduces the MiniMax M2—a 10-billion active parameter open-weight model designed for coding, agentic tasks, and long-horizon reasoning. Rather than leading with benchmark tables, the talk focuses on the training methodology behind M2’s real-world performance, explaining two distinctive approaches: scaled RL environments with verifiable coding goals, and in-house expert developers serving directly as reward models during training.
The expert developer feedback loop is a notable differentiator. MiniMax’s in-house engineers participated in defining training problems, evaluating model outputs for reliability, and providing precise reward signals around behaviors developers actually value—producing a model calibrated to practical development workflows rather than benchmark optimization alone. M2 also introduces interleaved thinking: an inference architecture where the model alternates between tool calls and reasoning steps across potentially tens to hundreds of turns within a single user interaction, enabling adaptation to noisy, dynamic real-world environments rather than committing to a single reasoning pass.
Within its first week of release, M2 reached the highest download count on Hugging Face for new models and climbed to top three in token usage on OpenRouter, suggesting strong early community adoption. The model supports full-stack multilingual coding, multi-agent coordination, and long-horizon automation tasks spanning tools like Gmail, Notion, and terminal environments—positioning it as a practical open-weight alternative for developers building agentic systems.
📺 Source: AI Engineer · Published December 13, 2025
🏷️ Format: Keynote Launch
