MiniMax has released and open-sourced MiniMax-M2, a compact Mixture of Experts (MoE) model designed for efficient coding and agentic tasks. The model features 230 billion total parameters with only 10 billion active parameters, aiming to provide elite performance in code generation and tool use while maintaining lower latency and cost.
- MiniMax-M2 ranks #1 among open-source models globally on the Artificial Analysis composite score for general intelligence.
- It excels at multi-file edits, coding-run-fix loops, and test-validated repairs, showing strong performance on Terminal-Bench and (Multi-)SWE-Bench.
- The model plans and executes complex toolchains across shell, browser, and code runners, with consistent results in BrowseComp-style evaluations.
- The 10 billion activated parameter design allows for faster feedback cycles, more concurrent runs, and simpler capacity planning for interactive agents.
MiniMax-M2 is available as open-source weights on Hugging Face and via the MiniMax Open Platform API, offering a deployment-friendly footprint for developers needing frontier-style capabilities without frontier-scale costs.