MiniMax releases MiniMax M3
MiniMax officially releases MiniMax M3 on June 1, 2026, an open-weight, natively multimodal model with a one-million-token context window.
What happened
Event details
MiniMax says M3 accepts image and video input and uses MiniMax Sparse Attention to support up to one million tokens of context. Its model card describes an image-text-to-text model with about 428B total parameters, about 23B active parameters and an open-weight license.
Assessment
Why it matters
The model combines long context, an MoE architecture, and image and video understanding in publicly available weights, giving it strong technical and ecosystem impact for long-horizon agents and multimodal development. Official materials do not indicate a major policy or social event.
Availability
Access notes
The official launch page provides API and agent access, while the model card includes downloads and local deployment instructions.