Summary

MiniMax-M2 is an open-source Mixture-of-Experts (MoE) model engineered for efficient and high-performance coding and agentic workflows. Targeting developers and researchers, it offers a streamlined form factor with 10 billion active parameters, delivering competitive general intelligence and advanced tool-use capabilities at lower latency and cost. This model aims to redefine efficiency for AI agents, making deployment and scaling more accessible without compromising on sophisticated capabilities.

How It Works

This MoE model features 230 billion total parameters but activates only 10 billion per inference, enabling a highly efficient "plan → act → verify" loop for agents. This design choice significantly reduces compute overhead, leading to faster feedback cycles in development tasks and more concurrent agent runs within budget constraints. Its architecture is optimized for sophisticated end-to-end tool use across various domains like shell, browser, and code execution, providing powerful capabilities in a compact, deployable package.

Quick Start & Requirements

Model weights are available on HuggingFace (MiniMaxAI/MiniMax-M2). Recommended inference frameworks include SGLang and vLLM, both offering day-0 support and deployment guides. The MiniMax Agent (agent.minimax.io) and MiniMax Open Platform API (platform.minimax.io) are also accessible, currently free for a limited time. Recommended inference parameters are temperature=1.0, top_p = 0.95, top_k = 40. Local deployment requires downloading weights and using compatible inference servers.

Highlighted Details

• Achieves the #1 composite score among open-source models globally according to Artificial Analysis benchmarks, demonstrating competitive general intelligence.
• Demonstrates strong performance in coding tasks, including multi-file edits and coding-run-fix loops, validated on SWE-Bench and Terminal-Bench.
• Excels in agentic workflows, planning and executing complex toolchains and effectively retrieving information in browsing tasks (BrowseComp).
• Its efficient 10B active parameter design ensures responsive agent loops and better unit economics for deployment.

Maintenance & Community

The project encourages feedback from developers and researchers. Contact is available via model@minimax.io. Collaborations with inference framework teams (SGLang, vLLM) are noted, indicating active ecosystem integration.

Licensing & Compatibility

The provided README does not explicitly state the license type or any compatibility notes for commercial use or closed-source linking.

Limitations & Caveats

Optimal performance requires retaining the <think>...</think> tags within historical messages, as the model is interleaved. The specific license for commercial use or integration into closed-source projects is not detailed in the README, which may pose an adoption blocker.