MiniMax-M3
MiniMax-M3
MiniMaxAI/MiniMax-M3 is MiniMaxAI’s 428B A22B Mixture-of-Experts first vision-language model combining long-context reasoning, agentic workflows, and creative capabilities in a single platform.
:::{card}
Positioning
MiniMax-M3 targets advanced use cases such as long-form video understanding, extended coding tasks (8+ hours), and high-quality design workflows supporting upto 1M context length.
Architecture
- Language backbone: derived from MiniMax-M2.7 with 60 layers (3 dense + 57 MoE), 64 attention heads, 128 experts, block-sparse attention called MiniMax Sparse Attention (MSA) on the MoE layers and a 512k context length.
- Vision backbone: CLIP-style ViT with 32 layers and dynamic resolution image input from 336×336 up to 2016×2016.
- Optimization target: trained on Hopper GPUs, with BF16 and FP8 support.
Key Strengths
- Ultra-long context via MSA (MiniMax Sparse Attention). A block-sparse attention design that avoids full attention’s quadratic cost enabling up to 1M-token context.
- Frontier coding and agentic performance. Positioned as a coding and agent-first model, with strong benchmark numbers and good at multi-turn, collaborative agent behavior.
- Native multimodality (image, video, desktop use). It supports image and video input and can operate a desktop computer.
- Long-horizon autonomous execution on complex real-world tasks. Can work over hours to a full day without human intervention.
Available Models
- MiniMax-M3 — registered as
MiniMaxM3SparseForConditionalGenerationmodel class.
Example HF Models
Example Recipes
- Full SFT — MedPix, EP32 + PP4
- Full SFT — MedPix CP2 comparison
- Full SFT — Tulu 3 text, CP8
- LoRA — MedPix, PP4 + EP8
See the MiniMax-M3 fine-tuning guide for the training setup and launch notes.