Chinese AI firm MiniMax has announced that its new AI reasoning model, MiniMax-M1, outperforms its domestic rival DeepSeek’s latest offering, DeepSeek-R1-0528, especially in efficiency and handling complex scenarios.
MiniMax claims its M1 model requires significantly less computing power for reasoning tasks. Specifically, it states that for deep reasoning with 80,000 tokens, M1 needs only about 30% of the computing power of DeepSeek R1. This is attributed to its “Lightning Attention” mechanism, which improves both training and inference efficiency.
M1 boasts a context window of one million tokens, which is eight times the capacity of DeepSeek R1 and rivals Google’s Gemini 2.5 Pro. This allows the AI system to process and understand significantly more information simultaneously.
The improved computational efficiency, combined with an advanced reinforcement learning algorithm called CISPO, results in lower training costs. MiniMax stated that the entire reinforcement learning phase for M1 used only 512 Nvidia H800 GPUs for three weeks, costing approximately $537,400. It is a huge achievement compared to the cost of $5.6 million on DeepSeek-R1-0528’s learning phase
MiniMax-M1 has been released under an Apache software license, making it open-source, unlike some other models that use more restrictive community licenses. The company asserts that M1’s capabilities are “top-tier among open-source models, surpassing domestic closed-source models and approaching the leading overseas models.” It claims M1 is competitive with models like OpenAI o3, Gemini 2.5 Pro, and Claude 4 Opus on various benchmarks (including AIME 2024, LiveCodeBench, SWE-bench Verified, Tau-bench, and MRCR).
MiniMax’s announcement underscores the fierce competition among Chinese AI startups. DeepSeek, which had previously disrupted the industry with its efficient and powerful R1 model, is now facing a challenge from a domestic peer just months after its own release.
MiniMax is backed by major Chinese tech companies like Alibaba Group and Tencent, as well as IDG Capital, indicating significant investment in the country’s AI development. The open-sourcing of models by Chinese firms like MiniMax and DeepSeek is a strategic move to democratize AI and potentially undercut the business models of Western proprietary labs. This approach aims to accelerate global adoption of their technology.