MiniMax M2 Explained

MiniMax M2 is an open-weight Mixture-of-Experts language model from MiniMax, a Chinese AI lab that is publicly listed on the Hong Kong Stock Exchange and backed by Alibaba and Tencent.

The flagship M2 checkpoint activates only 9.8 billion of its 229.9 billion total parameters per token, and MiniMax released it and the follow-up M2.5 under the permissive MIT license, so both are free to download and self-host, including for commercial use.

A newer M2.7 checkpoint changes the license: it stays open for personal use and non-commercial research, but any commercial deployment now requires prior written authorization from MiniMax.

This guide covers the architecture, what each licensed version actually allows, and the same data-sovereignty question that applies to every China-origin model: the hosted API runs in China, so self-hosting is the practical path for regulated data.

What Is MiniMax M2?

MiniMax M2 is an open-weight, decoder-only Mixture-of-Experts language model from MiniMax, released with weights publicly available on GitHub and Hugging Face.

The model uses a sparse Mixture-of-Experts design: 229.9 billion total parameters across 62 transformer layers, but only 9.8 billion parameters activate for any given token, which keeps inference cost down relative to a dense model of similar total size.

MiniMax, the developer, is a Chinese AI lab that is publicly listed on the Hong Kong Stock Exchange, with Alibaba and Tencent among its backers.

Architecture: How MiniMax M2 Is Built

Each of M2's 62 layers routes tokens across 256 fine-grained experts per Mixture-of-Experts layer, selecting 8 of them per token through sigmoid gating rather than the more common softmax routing.

Attention uses a grouped-query design with 48 query heads and 8 key-value heads per layer, plus QK-Norm and partial rotary position embeddings (RoPE). Notably, MiniMax tested a hybrid sliding-window attention setup during development and ultimately kept full attention in every layer instead.

The model was pretrained on 29.2 trillion tokens with a maximum training context length of 192K tokens, and it natively supports a context window of roughly 196,608 tokens (commonly rounded to 192K).

MiniMax trained the model specifically on executable software-engineering tasks and folded tool-call latency into its reinforcement-learning process, which is why the model is positioned for agentic and coding workloads rather than as a general chat assistant.

M2 vs M2.5 vs M2.7: Which License Applies

MiniMax has shipped three versions under two different license terms, and the version you pick determines whether commercial self-hosting is straightforward or requires MiniMax's sign-off.

M2 and M2.5 ship under the plain MIT license: you can download the weights, self-host them, modify them, and use them commercially with no revenue thresholds and no attribution requirement.

M2.7 ships under a Modified-MIT license instead. Personal use, self-hosting, and non-commercial research or education stay free and unrestricted. Commercial use — including commercially self-hosting the weights, using MiniMax's own paid API commercially, or deploying a fine-tuned derivative for a paid product — requires prior written authorization from MiniMax, requested through api@minimax.io, and a "Built with MiniMax M2.7" attribution notice once authorized.

If you want a commercial open-weight model you can self-host today with no extra licensing step, M2 or M2.5 is the safer pick. M2.7 is worth watching for its architecture improvements, but budget time for the authorization step before you build a commercial product on it.

Data Sovereignty and Compliance

MiniMax is a China-based lab, and its hosted API runs on infrastructure inside that jurisdiction. As with every model in our Chinese AI models roundup, sending regulated data to the hosted API is an offshore data transfer under Chinese data law.

Self-hosting removes that specific exposure. Because M2 and M2.5 are plain MIT, a regulated US or EU business can download the weights and run them entirely on infrastructure it controls, with no separate agreement needed.

MiniMax does not publish a BAA, SOC 2 report, or HIPAA attestation for its own hosted API. Treat the hosted API the same way you would treat any other China-hosted model API: fine for evaluation, not a default choice for regulated production data.

How MiniMax M2 Compares

Versus DeepSeek-V3: DeepSeek-V3 is the larger model (671B total / 37B active parameters) and has a longer independent-benchmark track record. MiniMax M2 is more parameter-efficient per token (9.8B active) and is purpose-trained on agentic, tool-using coding tasks — pick DeepSeek for broad reasoning workloads with more third-party validation, MiniMax M2 for agent/tool-call-heavy pipelines.

Versus Qwen 3.6: Qwen 3.6 remains the stronger pick for multilingual work and ships under the equally permissive Apache 2.0 license. MiniMax M2's focus is narrower — coding and agentic tool use — so the choice mostly comes down to workload fit rather than licensing.

Versus Claude Opus 5: Claude Opus 5 is a closed, hosted-only model with a signed BAA available for regulated US customers; MiniMax M2 is open-weight and self-hostable but carries the China-origin data question. See our dedicated Claude Opus 5 vs MiniMax M2 comparison for the full breakdown.

Selection principle: for an agentic/coding pipeline you plan to self-host, MiniMax M2 is a legitimate open-weight option to shortlist alongside DeepSeek and Qwen — evaluate all three on your own tasks rather than on parameter count alone.

How to Access and Deploy MiniMax M2

Self-hosted via open weights: download M2 or M2.5 from GitHub or Hugging Face and run them with standard inference frameworks (vLLM, SGLang, and similar MoE-aware runtimes). The sparse 9.8B-active-parameter design keeps inference cheaper than a dense model of the same total size, but 229.9B total parameters still means substantial accelerator memory for full precision — plan for quantization if your hardware is limited.

Hosted API: MiniMax also offers its own hosted API through its developer platform, billed per token. Confirm current rates directly on MiniMax's own pricing page before budgeting — we do not quote a figure here because publicly reported numbers for this API varied significantly across sources at the time of writing.

M2.7 commercial deployment: if you want to commercially self-host or productize M2.7 specifically, request authorization from MiniMax at api@minimax.io before you build past a prototype.

Is MiniMax M2 Ready for Business Use?

Best fit: internal agentic/coding tooling, development environments, and self-hosted automations where the sparse MoE design's cost efficiency matters and you control the infrastructure.

License clarity matters more here than with most open models — confirm whether you are deploying M2/M2.5 (plain MIT, no extra step) or M2.7 (commercial use requires MiniMax's written authorization) before you scale past a pilot.

As with every model in our Chinese AI models roundup, keep regulated or customer data off the hosted API and self-host on infrastructure you control if compliance is a requirement.