MiniMaxAI/MiniMax-M3

MiniMax-M3: A Native Multimodal Model

MiniMax-M3, developed by MiniMaxAI, is a powerful native multimodal model with approximately 428 billion parameters and 23 billion activated parameters, supporting an extensive 1 million token context length. Unlike models that add modalities later, M3 undergoes mixed-modality training from its initial stages, enabling deeper and more integrated semantic understanding across text, image, and video.

Key Capabilities & Innovations

• Native Multimodality: Achieves profound semantic fusion by training on diverse modalities (text, image, video) from the outset.
• Context Scaling with MiniMax Sparse Attention (MSA): Integrates MSA, a high-performance sparse attention operator, to dramatically improve efficiency for 1M token contexts. This results in 9x prefill and 15x decode speedups compared to M2, reducing per-token compute by 20x.
• Coding & Cowork: Demonstrates frontier-level performance in long-horizon agentic tasks, particularly strong in coding and collaborative work scenarios.
• Flexible Reasoning Modes: Offers enabled, adaptive, and disabled reasoning modes via the thinking parameter, allowing users to balance reasoning depth with latency and throughput requirements.

Good For

• Applications requiring deep multimodal understanding and generation.
• Tasks benefiting from extremely long context windows, such as complex document analysis or extended conversations.
• Agentic workflows, coding assistance, and collaborative development environments.
• Developers seeking efficient inference for large multimodal models through optimized attention mechanisms.