MiniMax · open weights · updated July 2026

MiniMax-M2

Parameters
230B / 10B active
Context window
~200K tokens
License
MIT
Modalities
text
Released
2025-10-27

MiniMax-M2 is the tool-use specialist of the open-weight field: 230B total parameters with just 10B active, MIT-licensed, and tuned specifically for agentic loops — 69.4% SWE-bench Verified, 46.3 Terminal-Bench, 83 LiveCodeBench. Mid-2026 roundups still rank it the best open model for tool-driven agents.

The 10B active footprint is the strategic choice: agent workloads burn tokens on long multi-step loops, and M2 delivers near-frontier agentic quality at interactive speed and low serving cost. For teams building autonomous agents rather than chat products, M2 is frequently the efficiency-optimal pick.

MiniMax-M2 specifications

VendorMiniMax (China)
ArchitectureMixture-of-Experts, agent- and tool-use-optimized
Total parameters230B
Active parameters10B
Context window~200K tokens
Modalitiestext
LicenseMIT — weights free to download, self-host, fine-tune, and use commercially
Release date2025-10-27
WeightsHugging Face · GitHub

Benchmarks and reported results

BenchmarkResultNote
SWE-bench Verified69.4%
Terminal-Bench46.3
LiveCodeBench83

Running MiniMax-M2 locally

230B total but only 10B active: 4-bit builds run around 130GB, and the tiny active footprint makes it unusually fast per token. One of the best capability-per-dollar self-host targets for agent backends.

What MiniMax-M2 is best for

  • Tool-calling agents and terminal/computer-use workloads
  • High-throughput agent loops where per-token cost compounds
  • Self-hosting on a modest multi-GPU node (MIT license)

Frequently asked questions

What makes MiniMax-M2 good for agents specifically?

It was optimized for multi-step tool use rather than single-shot chat: strong Terminal-Bench (46.3) and SWE-bench Verified (69.4%) scores with only 10B active parameters, so long agentic loops stay fast and cheap.

Can I self-host MiniMax-M2?

Yes — MIT license, ~130GB at 4-bit for the 230B weights. A 2×80GB node or high-memory workstation handles it, with vLLM and SGLang support.