MiniMax-M3 is a multimodal foundation model developed by MiniMax, officially released in June 2026. It is designed to provide high-performance capabilities in coding, autonomous agentic reasoning, and long-context processing. Unlike many contemporary models that integrate vision through secondary fine-tuning, M3 is a natively multimodal system trained on interleaved text, image, and video data from its initial pretraining stages.
Architecture and Context
The model is built on the MiniMax Sparse Attention (MSA) architecture, which optimizes context scaling by utilizing KV-block selection instead of traditional full attention. This architectural innovation enables a 1-million-token context window while significantly reducing computational overhead and increasing decoding speeds compared to previous generations. The architecture is specifically tuned for long-horizon tasks, such as full-repository code analysis and multi-step autonomous workflows that require sustained coherence over extremely large data volumes.
Capabilities and Usage
In specialized benchmarks, MiniMax-M3 has demonstrated proficiency in software engineering and autonomous web navigation, achieving notable scores on evaluations like SWE-Bench Pro and BrowseComp. It is capable of independently managing complex projects, such as replicating research experiments and optimizing low-level CUDA kernels. The model's native multimodality also allows it to interpret charts, complex formulas, and video frames with high precision.
For effective prompting, users are encouraged to take advantage of the model's native support for multiple input types, including images and video URLs. When working with very large datasets, it is often recommended to provide only the relevant segments of a codebase or document to manage operational costs, as some implementations utilize tiered billing for requests exceeding 512,000 tokens.