MiniMax Releases Open-Source Reasoning Model
MiniMax has released MiniMax-M1, which it claims is the first open-source, large-scale reasoning model with a hybrid-attention mechanism. With open weights and a permissive license, it's aimed at accelerating research and development for custom agentic systems that require multi-step reasoning.
The model's architecture combines a Mixture-of-Experts (MoE) design with a "lightning attention" mechanism. This allows it to have a massive 456 billion total parameters, but only activate a fraction of them (45.9 billion) for any given task, leading to significant computational efficiency. This efficiency is particularly notable in long-context tasks. For a generation length of 100,000 tokens, MiniMax-M1 uses only 25% of the computational resources (FLOPs) compared to models like DeepSeek R1. The model supports a context window of up to 1 million tokens, which is 8 times larger than that of DeepSeek R1. On performance benchmarks, the MiniMax-M1-80k version demonstrates leading performance among open-weight models in software engineering, tool use, and long-context tasks. For example, it scored 56.0% on the SWE-bench validation benchmark and has shown strong results on mathematics and reasoning benchmarks like AIME and GPQA Diamond. The model was trained using a novel reinforcement learning algorithm developed by MiniMax called CISPO, which clips importance sampling weights to improve efficiency. The entire reinforcement learning phase for the model cost just over $500,000. MiniMax, the company behind the model, was founded in 2021 and is considered one of China's "Six AI Tigers." The Shanghai-based firm develops both foundation models and consumer applications, and it went public on the Hong Kong Stock Exchange in January 2026. In March 2024, Alibaba led a $600 million funding round that valued the company at $2.5 billion.
