MiniMax Open-Sources MiniMax-01 Series with Lightning Attention for Long Context
MiniMax has released its MiniMax-01 series, featuring the foundational language model MiniMax-Text-01 and the multimodal MiniMax-VL-01. These models introduce 'Lightning Attention,' an alternative to the traditional Transformer architecture, enabling them to process context lengths of up to 4 million tokens, significantly larger than other leading models. The series has a 456 billion parameter count, with 45.9 billion parameters activated per inference. MiniMax has open-sourced the complete weights of both models to encourage further research and claims their performance is on par with global leaders.
This development pushes the boundaries of AI agent capabilities by addressing the need for sustained memory and extensive inter-agent communication through ultra-long context windows, accelerating the AI Agent era. The open-sourcing allows broader adoption and contributions to long-context AI research.
Learn one new AI thing every day.
Daily Deck sends you seven plain-English cards like this every morning. Free.
Start free