MiniMax-M1 is a large-scale, open-weight reasoning model designed for extended context and high-efficiency inference. It leverages a hybrid Mixture-of-Experts (MoE) architecture paired with a custom "lightning attention" mechanism, allowing it to process long sequences—up to 1 million tokens—while maintaining competitive FLOP efficiency. With 456 billion total parameters and 45.9B active per token, this variant is optimized for complex, multi-step reasoning tasks. Trained via a custom reinforcement learning pipeline (CISPO), M1 excels in long-context understanding, software engineering, agentic tool use, and mathematical reasoning.
Evaluations
17
across 17 benchmarks
Latency
—
Context length
1M
tokens
Cost
$0.40 · $2.20
input · output per 1M tokens
Key takeaways
MiniMax M1 is an open-weight, large-scale reasoning model designed for extended context (up to 1 million tokens) and high-efficiency inference. It excels in complex, multi-step reasoning tasks such as software engineering, agentic tool use, and mathematical reasoning.
The model utilizes a hybrid Mixture-of-Experts (MoE) architecture with a custom 'lightning attention' mechanism. It has 456 billion total parameters (45.9B active per token) and was trained using a custom reinforcement learning pipeline (CISPO).
MiniMax M1 demonstrates strong capabilities in easy-level scientific reasoning, achieving 98.9% accuracy on the AI2 Reasoning Challenge - Easy, showcasing its reasoning strength and capacity for long-context understanding.
The model's significant limitation is its struggle with code generation and fixing tasks, achieving 0% accuracy on SWE-bench Lite evaluations. It often gets stuck in observational loops, failing to translate problem descriptions into actionable code fixes or targeted suggestions.
Evaluations show a stark contrast: 98.9% accuracy on easy-level scientific reasoning (AI2 Reasoning Challenge - Easy) versus 0% accuracy on code modification and bug-fixing tasks (SWE-bench Lite) due to repetitive exploration without generating solutions.
