MiniMax is an AI company whose language models, such as MiniMax-Text-01 and the open-weight MiniMax-M1, are known for very long context windows of around a million tokens.
In plain words
Normal attention gets much slower as the input grows. MiniMax models mix a faster method called lightning attention with a mixture-of-experts design: 456 billion parameters in total, about 46 billion used per token. MiniMax-Text-01, from January 2025, can train on a million-token context. MiniMax-M1, from June 2025, adds step-by-step reasoning trained with reinforcement learning, and its repository uses the Apache 2.0 licence. MiniMax also makes the Hailuo video models.
A quick example
Imagine handing a model an entire long codebase or several books at once and asking where a certain idea first appears. A model with a million-token context can read all of it in one go instead of in pieces.
Where you'll meet it
- Tasks with very long inputs, such as whole documents or codebases.
- Reasoning and software engineering agents.
Related
Sources
- PaperMiniMax-01: Scaling Foundation Models with Lightning Attention, MiniMax (arXiv) · read 29 Sept 2026
- PaperMiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention, MiniMax (arXiv) · read 29 Sept 2026
- DocsMiniMax-M1, MiniMax (GitHub) · read 29 Sept 2026
- DocsMiniMax Hailuo 02, World-Class Quality, Record-Breaking Cost Efficiency, MiniMax · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.