MiniMax M3 is presented as the first open-weight model to simultaneously achieve three frontier capabilities: advanced coding and agentic capabilities, a 1-million-token context window powered by the MiniMax Sparse Attention (MSA) architecture, and native multimodality.

Core Methodology and Capabilities:

1. MiniMax Sparse Attention (MSA) and Extended Context Window:

The M3 model leverages a proprietary MiniMax Sparse Attention (MSA) architecture to achieve an extensive context window. While the specific mechanism of MSA is not detailed, it is stated to be the underlying infrastructure that enables an input context window of up to 1 million tokens, with a guaranteed minimum of 512,000 tokens. This substantial context capacity is crucial for handling long-range tasks in domains such as complex multi-step agentic workflows, long-form codebases, and comprehensive video understanding.

2. Native Multimodality:

Multimodality in M3 is a core, integrated capability rather than an superficial add-on. The entire data pipeline for M3 was re-engineered and scaled to include over 100 trillion ($100T+$) tokens of pretraining data. Crucially, multimodal training was integrated "from step zero," ensuring deep alignment between textual and visual semantic spaces. This deep integration allows the model to process and understand information across different modalities inherently. Its proficiency in multimodal tasks is demonstrated on the BrowseComp benchmark, where M3 scores 83.5, outperforming Opus 4.7's 79.3, indicating strong autonomous browsing and information retrieval abilities.

3. Frontier Coding and Agentic Capabilities:

M3 exhibits top-tier performance in coding and agentic benchmarks. Its capabilities include autonomous task decomposition, sophisticated tool invocation, and multi-step reasoning. These features provide a robust foundation for building AI coding assistants and automating complex workflows.

Performance and Demonstrations:

The model's advanced capabilities are validated through several long-horizon, autonomous demonstrations:

• ICLR Paper Reproduction (12-Hour Autonomous Replication): M3 was tasked with independently replicating an ICLR 2025 paper, "Learning Dynamics of LLM Finetuning." Over approximately 12 hours, M3 operated continuously, autonomously generating 18 Git commits and 23 experimental figures, successfully replicating the core experiments. This task showcased its multimodal ability to parse charts and formulas from the paper, its long context capacity to manage the paper, associated code, and experiment logs within a single context window, and its coding/agentic capabilities to drive the long-horizon execution.

• CUDA Kernel Optimization (9.4x Speedup): M3 demonstrated its coding and optimization prowess by optimizing an FP8 GEMM (General Matrix Multiply) kernel, a compute-intensive operation for LLM inference, on NVIDIA Hopper GPUs. Starting with only a task description and a non-runnable Triton skeleton, M3 executed 147 benchmark submissions and 1,959 tool calls over approximately 24 hours. This autonomous process resulted in an increase in hardware peak utilization from 7.6% to 71.3%, achieving a 9.4x speedup with zero human intervention.

• PostTrainBench (Autonomous Model Training): M3 was given four pretrain-only base models and tasked with autonomously completing the full pipeline of data synthesis, training, evaluation, and iteration within 12 hours, aiming to enable these models to perform mathematical reasoning, code generation, and knowledge Q&A. The entire process ran without human intervention. M3 achieved a score of 37.1 on this benchmark, ranking third overall, behind Opus 4.7 (42.4) and GPT-5.5 (39.3), significantly surpassing other models.

Developer Access and Future Plans:

MiniMax M3 is accessible via an API at https://api.minimax.io/v1/text/chatcompletion_v2 using the MiniMax-M3 model identifier. The platform supports various integrations, including standard API access for developers, enhancements for existing token plan users, and integration with the MiniMax Code platform for experiencing agentic and multimodal capabilities without development. Furthermore, MiniMax has committed to open-sourcing M3 on HuggingFace and GitHub, promising support for private cluster deployment and fine-tuning, thereby contributing to the open-source community.