The Chinese startup MiniMax, like their colleagues from DeepSeek, decided to host an actual release week. They already have something to show:
On Monday, the world saw the first reasoning model of the team — M1. It is open-source and features a very large context window of one million tokens. The longest processing window in the world. It also has agentic capabilities, making it more versatile. During training, they managed to spend just around $500,000, and in tests, it slightly trails behind Gemini 2.5 Pro. Reports, code, and model weights are available to the public.
Yesterday, a new model, Hailuo 2, was introduced, combining text and image processing for video creation. It works exceptionally well with physical models and complex movements, and demonstrates a good understanding of instructions. Developers emphasize that it has record-breaking performance and is very affordably priced. You can try it for free right now.
Overall, the launch week for MiniMax was intense and substantial. More interesting announcements are expected today!