China’s MiniMax Unveils AI Model M1
Shanghai-based startup MiniMax has unveiled a new language model with step-by-step reasoning, M1. The developers claim that it consumes half as much computing power as the popular DeepSeek-R1, while maintaining comparable accuracy.
According to the developers, this algorithm can compete with similar ones from leading industry players, including OpenAI, Anthropic, and DeepSeek, in terms of performance and creativity, but at the same time, the new AI model is significantly cheaper to train and operate. MiniMax said it spent only $534,700 on renting data center computing power to train M1. By comparison, industry experts estimate that training the GPT-4o algorithm cost about 200 times more, and OpenAI spent more than $100 million on it. The developer has not disclosed official data on this issue.
The advantage of M1 is its extended context window: up to 1 million input tokens and up to 80 thousand output ones. This makes the model suitable for tasks related to processing large volumes of information: from corporate documentation to analytical reports.
MiniMax is reportedly backed by some of China’s largest tech companies, including Tencent and Alibaba. It’s unclear how many people work at the company, and there’s little information about MiniMax CEO Yan Junjie.
