MiniMax M2.7

MiniMax M2.7 is a large language model (LLM) developed by the Chinese company MiniMax Group Inc. (稀宇科技), released on March 18, 2026. It is a Mixture of Experts (MoE) model with 230 billion total parameters, of which only 10 billion are activated during a single inference pass. M2.7 is the direct successor to MiniMax M2.5 (February 2026) and represents the first iteration in the M2 series in which the model actively participated in its own evolution and refinement process.

Architecture and parameters

The model is based on a sparse Mixture of Experts architecture with 256 experts and a top-k routing mechanism. It employs multi-head causal self-attention augmented with Rotary Position Embeddings (RoPE) and Query-Key Root Mean Square Normalization (QK RMSNorm). The context window is 200,000 tokens, and the maximum output length reaches 131,072 tokens. By activating only ~10B parameters per forward pass, the model offers low latency and reduced inference costs while maintaining capabilities comparable to dense models.

Self-directed model evolution

A key distinguishing feature of M2.7 is that an internal version of the model actively participated in its own training process. In this experiment, the model autonomously optimized a programming scaffold over more than 100 rounds — analyzing failure trajectories, modifying code, running evaluations, and deciding whether to accept or revert changes — achieving a 30% performance improvement without human intervention. The model constructs complex agentic harnesses, updates its own memory, creates dozens of compound skills, and refines its own learning process based on experimental results.

Software engineering and agentic workflows

M2.7 is designed primarily for advanced programming tasks and long agentic chains. It supports log analysis, bug detection, refactoring, code security, machine learning tasks, and comprehensive end-to-end project delivery. The model natively supports Agent Teams (multi-agent collaboration) and dynamic tool retrieval. On the SWE-Pro benchmark it scored 56.22%, matching GPT-5.3-Codex. On Terminal Bench 2 it achieved 57.0%, on VIBE-Pro — 55.6%, on SWE Multilingual — 76.5, and on MLE Bench Lite a medal rate of 66.6% (second among open-weight models).

Professional office work

The model demonstrates strong capabilities in office document editing — Excel, Word, and PowerPoint — with support for multi-round, high-fidelity modifications. On the GDPval-AA benchmark it achieved an ELO score of 1495, the highest among open-weight models. On Toolathon it reached 46.3% accuracy, and on MM Claw — 62.7%, approaching Sonnet 4.6. The model maintains a 97% skill compliance rate across a set of 40 complex skills each exceeding 2,000 tokens.

Availability and license

Model weights are publicly available on Hugging Face (MiniMaxAI/MiniMax-M2.7) and in the GitHub repository (MiniMax-AI/MiniMax-M2.7). The model can be run locally using SGLang, vLLM, or Transformers, as well as through the NVIDIA NIM Endpoint and Ollama. The license, described by MiniMax as "Modified-MIT," is in practice a non-commercial license — non-commercial use is permitted, while any commercial use requires prior written consent from MiniMax. M2.7 is the first model in the M2 series to depart from the fully permissive MIT license used in earlier versions (M2, M2.1, M2.5).

API pricing

Under the Pay-as-You-Go model, inference via the MiniMax API costs $0.30 per million input tokens and $1.20 per million output tokens. A MiniMax-M2.7-highspeed variant is also available ($0.60/$2.40 per million tokens), offering higher throughput at the same quality. Prompt caching read costs $0.06/M tokens, and prompt caching write — $0.375/M tokens.