MiniMax M3

MiniMax M3 is the newest entry in MiniMax's "M" series of text/reasoning models, positioned by the company as a frontier model for coding, agentic workflows, and complex reasoning. According to MiniMax, M3 reaches frontier capability on coding and agentic tasks, introduces a new MiniMax Sparse Attention (MSA) mechanism supporting up to a one-million-token context, and is natively multimodal. Its API supports text, image, and video input through OpenAI- and Anthropic-compatible interfaces.

The headline architectural change versus its predecessors is MSA. Where the prior [[sibling:minimax-m27|MiniMax M2.7]] was a mixture-of-experts model with 230 billion total parameters, 10 billion active per token, 256 experts, and a roughly 200K-token context, M3's sparse-attention design is the provider's reported route to handling far longer documents, codebases, and multi-step agent sessions more efficiently. MiniMax presents M3 as the successor while keeping M2.7 and earlier models available for existing pipelines.

For context on the lineage, MiniMax reported that [[sibling:minimax-m25|MiniMax M2.5]] scored 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp (with context management), and completed SWE-Bench Verified evaluation 37% faster than M2.1. MiniMax has not published comparable official M3 benchmark figures here, so specific scores are omitted pending the model card.

Note that MiniMax's documentation cites the one-million-token figure for M3, whereas this catalog entry lists a 198,000-token window; treat the provider's specification as authoritative and verify the exact limit at launch.

Sources

Model Invocation - MiniMax API Docs· platform.minimax.ioMiniMax M2.7 Advances Scalable Agentic Workflows on NVIDIA Platforms for Complex AI Applications | NVIDIA Technical Blog· developer.nvidia.comMiniMax· minimax.iominimax-m2.7 Model by Minimaxai· build.nvidia.comMiniMaxAI/MiniMax-M2.5 · Hugging Face· huggingface.co

This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.