Chinese AI Model MiniMax-M1 Surpasses DeepSeek in Efficiency
Chinese tech company MiniMax has introduced its new AI reasoning model called MiniMax-M1, which it says is better than its local competitor DeepSeek’s latest model, DeepSeek-R1-0528. MiniMax claims that M1 is more efficient and handles complex tasks better than DeepSeek-R1.
According to the company, M1 uses much less computing power for reasoning. For example, when working with 80,000 tokens, it only needs about 30% of the computing power that DeepSeek R1 requires. This is because of a special system called “Lightning Attention” that makes both training and task performance faster and more efficient.
MiniMax-M1 also has a much larger memory capacity. It can handle up to one million tokens at a time, which is eight times more than DeepSeek R1 and on the same level as Google’s Gemini 2.5 Pro. This allows the model to process a lot more information at once.
Another reason M1 is more efficient is due to a new reinforcement learning method called CISPO. Because of this, training the model cost much less. MiniMax said that training M1 only needed 512 Nvidia H800 GPUs over three weeks, which cost around $537,400. In comparison, DeepSeek-R1-0528’s training phase cost about $5.6 million.
The M1 model is also open-source under the Apache license, meaning anyone can use or improve it. This is different from some other AI models that are kept private or restricted. MiniMax says that M1 performs at a top level among all open-source models and even beats some closed-source models in China. It claims M1 can compete with major global models like OpenAI’s o3, Gemini 2.5 Pro, and Claude 4 Opus in different benchmark tests like AIME 2024, LiveCodeBench, SWE-bench Verified, Tau-bench, and MRCR.
MiniMax’s launch of M1 highlights the growing competition between AI startups in China. DeepSeek, which had gained attention with its powerful R1 model, now faces strong competition just a few months after its release.
MiniMax is supported by big Chinese companies like Alibaba, Tencent, and IDG Capital. The move to release models as open-source, as MiniMax and DeepSeek have done, is seen as a way to make AI more widely available and possibly challenge the business models of Western AI companies. This strategy is also expected to help Chinese AI technology spread more quickly around the world.
