Meet MiniMax's New Open-Source LLM: A Game-Changer for AI Agents

Ever heard of MiniMax? They're not just about creating realistic AI videos like Hailuo. Today, they've unveiled something huge: the MiniMax-01 series. This includes MiniMax-Text-01, a large language model (LLM), and MiniMax-VL-01, a visual multi-modal model. The star of the show is MiniMax-Text-01, which can handle a massive 4 million tokens in its context window. That's like reading a small library in one go! Google's Gemini model could only manage 2 million tokens, but MiniMax doubled that. These models are available for free on platforms like Hugging Face and GitHub, under a custom MiniMax license. You can try them out on Hailuo AI Chat, which competes with big names like ChatGPT. The pricing is competitive too. For instance, MiniMax's API costs $0. 2 per 1 million input tokens and $1. 1 per 1 million output tokens, which is much cheaper than OpenAI's GPT-4. The secret behind MiniMax-01 is its Lightning Attention mechanism. Unlike traditional transformer architectures, it uses a mix of linear and SoftMax layers. SoftMax layers turn numbers into probabilities to figure out the most likely meaning of the input. This design makes the model more efficient and reduces computational complexity. On standard text and multi-modal benchmarks, MiniMax-01 performs as well as top models like GPT-4. It's especially strong on long-context evaluations. For example, MiniMax-Text-01 got 100% accuracy on the Needle-In-A-Haystack task with a 4-million-token context. MiniMax plans to keep improving these models, adding code and multi-modal enhancements. They believe open-sourcing will help build the foundation for AI agents, which are set to become more important in 2025. With its focus on cost-effective and scalable AI, MiniMax is ready to shape the future of AI agents.