Singapore-based AI company MiniMax has launched a new open-source language model that can process up to 4 million tokens at once, doubling the previous record. According to Carl Franzen’s report in VentureBeat, the MiniMax-01 series includes both text and visual capabilities. The model uses an innovative “Lightning Attention” architecture and mixture of experts framework with 32 experts to optimize performance. MiniMax offers API access at $0.2 per million input tokens, significantly lower than competitors like OpenAI. The company claims their model matches top performers like GPT-4 on standard benchmarks while achieving 100% accuracy on long-context tasks. The technology is now available through Hugging Face and Github, with MiniMax actively seeking collaboration from developers and researchers.