06.07.2025 06:41

MiniMax Unveils M1 Language Model, Outpacing DeepSeek R1 and Chinese Closed-Source Models

News image

MiniMax, a rising Chinese AI startup, has launched its new language model, M1, claiming it surpasses all closed-source Chinese models and outperforms DeepSeek R1 in several benchmarks.

The model boasts an impressive context window of one million tokens — eight times that of DeepSeek R1 — and in some cases requires only 30% of the computational resources of its competitor, marking a significant leap in efficiency.

The training of M1 utilized 512 Nvidia H800 GPUs, with a rental cost of $534,700. The model features a total of 456 billion parameters, though only 45.9 billion are activated per token, optimizing its performance for complex tasks.

M1 is available on HuggingFace in two reasoning budget variants, catering to diverse computational needs.

Also read:


In addition, MiniMax has introduced an upgraded version of its video model, Hailou O2. The initial release was already well-received, and the new iteration promises enhanced 1080p resolution along with improved accuracy in rendering complex physical details, further solidifying MiniMax’s position in the AI landscape.


0 comments
Read more