MiniMax Open-Sources Large-Scale Hybrid Architecture Reasoning Model | Counselor Vitality

MiniMax has announced the launch of MiniMax-M1, the world's first open-source large-scale hybrid-architecture reasoning model.

According to the company, M1 currently ranks in the top tier among open-source models for complex productivity-oriented scenarios, surpassing domestic closed-source models and approaching the most advanced overseas models.

Long-context capability is M1's standout strength. The model currently supports 1 million tokens of context input — eight times that of DeepSeek R1 and on par with Google Gemini 2.5 Pro. Its reasoning output ceiling reaches 80,000 tokens, the longest in the industry.

This advantage stems primarily from two novel proprietary technologies.

First, a hybrid architecture centered on Lightning Attention. Even for deep reasoning at a scale of 80,000 tokens total, it requires only about 30% of the compute power needed by DeepSeek R1 under equivalent conditions, improving efficiency during both training and inference.

The second innovation is the reinforcement learning algorithm CISPO, which dramatically boosts RL efficiency by clipping importance sampling weights rather than using conventional token-level updates.

By the company's estimates, the entire reinforcement learning phase took just three weeks, used 512 H800 GPUs, and cost only $537,400 (approximately RMB 3.839 million) in rental fees — far below projected spending levels.

Detailed benchmark results for M1 across 17 mainstream industry evaluation sets

The data above shows:

• MiniMax-M1-40k and MiniMax-M1-80k achieved 55.6% and 56.0% respectively on the SWE-bench Verified benchmark, slightly trailing DeepSeek-R1-0528's 57.6% but significantly outperforming all other open-weight models.
• Leveraging its million-token context window, the M1 series excels at long-context understanding tasks, surpassing not only all open-weight models but also OpenAI o3 and Claude 4 Opus, ranking second globally with only a narrow gap behind Gemini 2.5 Pro.
• In agentic tool-use scenarios (TAU-bench), MiniMax-M1-40k likewise leads all open-weight models.

M1 is currently available with unlimited free access on the MiniMax app and web, and API access is offered at the industry's lowest prices on the company's official website.

Over the next four days, MiniMax will roll out additional updates — stay tuned.