MiniMax, a Shanghai-based Chinese AI company, has recently released MiniMax-M1, the world’s first open-weight, large-scale hybrid-attention reasoning model. This model represents a significant breakthrough in AI reasoning capabilities and efficiency.
- Scale and Architecture: MiniMax-M1 is built on a massive 456 billion parameter foundation, with 45.9 billion parameters activated per token. It employs a hybrid Mixture-of-Experts (MoE) architecture combined with a novel “lightning attention” mechanism that replaces traditional softmax attention in many transformer blocks. This design enables the model to efficiently handle very long contexts—up to 1 million tokens, which is 8 times longer than the context size of DeepSeek R1, a leading competitor.
- Efficiency: The lightning attention mechanism significantly reduces computational cost during inference. For example, MiniMax-M1 consumes only 25% of the floating-point operations (FLOPs) required by DeepSeek R1 when generating 100,000 tokens, making it much more efficient for long-context reasoning tasks.
- Training and Reinforcement Learning: The model was trained using large-scale reinforcement learning (RL) on diverse tasks ranging from mathematical reasoning to complex software engineering environments. MiniMax introduced a novel RL algorithm called CISPO, which clips importance sampling weights rather than token updates, improving training stability and performance. Two versions of the model were trained with thinking budgets of 40,000 and 80,000 tokens respectively
Performance and Benchmarking
MiniMax-M1 outperforms other strong open-weight models such as DeepSeek-R1 and Qwen3-235B across a variety of challenging benchmarks including:
- Extended mathematical reasoning (e.g., AIME 2024 and 2025)
- General coding and software engineering tasks
- Long-context understanding benchmarks (handling up to 1 million tokens)
- Agentic tool use tasks
- Reasoning and knowledge benchmarks
For instance, on the AIME 2024 math benchmark, the MiniMax-M1-80K model scored 86.0%, competitive with or surpassing other top models. It also shows superior performance in long-context tasks and software engineering benchmarks compared to DeepSeek and other commercial models
Strategic and Industry Impact
MiniMax-M1 is positioned as a next-generation reasoning AI model that challenges the dominance of DeepSeek, a leading Chinese reasoning-capable large language model. MiniMax’s innovation highlights the rapid advancement and growing sophistication of China’s AI industry, especially in developing models capable of advanced cognitive functions like step-by-step logical reasoning and extensive contextual understanding.
The model’s release underscores the strategic importance China places on AI reasoning capabilities for applications across manufacturing, healthcare, finance, and military technology. MiniMax’s approach, combining large-scale hybrid architectures with efficient reinforcement learning and long-context processing, sets a new benchmark for open-weight models worldwide.
As a summary
- MiniMax-M1 is the world’s first open-weight, large-scale hybrid-attention reasoning model with 456 billion parameters.
- It supports extremely long context lengths (up to 1 million tokens) and is highly efficient, using only 25% of the compute of comparable models like DeepSeek R1 at long generation lengths.
- The model excels in complex reasoning tasks, software engineering, and tool use benchmarks.
- It is trained with a novel reinforcement learning algorithm (CISPO) that enhances training efficiency and stability.
- MiniMax-M1 represents a major step forward in China’s AI capabilities, challenging established players and advancing the global state of reasoning AI.