MiniMax M2

MiniMax M2 is an open-weight Mixture-of-Experts (MoE) language model released in late October 2025, purpose-built for coding and agentic workflows. With 230 billion total parameters but only 10 billion activated per token, it achieves a strong balance between capability and inference efficiency — the same core architecture carried forward into the M2.1, M2.5, and M2.7 successors. M2 debuted as the highest-performing open-weight model on Artificial Analysis' composite intelligence benchmark at launch, with particular strengths in SWE-bench style coding evaluations and long-horizon tool-use tasks across shell, browser, retrieval, and code-runner environments.

For a TheRouter operator, M2 occupies a unique position: it is the foundational release of the M2 family, demonstrating that a 10B-active-parameter MoE could compete with frontier models on coding and agent benchmarks. While its younger siblings (M2.1, M2.5, M2.7) have incrementally improved, M2 remains relevant as a well-documented baseline — it was the first open-weight model to prove that sub-20B-active architectures could deliver frontier-level SWE-bench scores (69.4% Verified), BrowseComp competitiveness (44.0%), and strong GAIA text-only performance (75.7%). New agents should default to M2.5 or M2.7, but M2 is useful for cross-version regression testing, academic baselines, and deployments that match a specific known M2 behavioural profile.