MiniMax训练推理模型仅花384万，售价对标豆包

Core Insights - MiniMax has launched its first inference model M1, which supports an industry-leading context input length of 1 million tokens, significantly surpassing DeepSeek R1's 125,000 tokens [2] - The M1 model utilizes a Mixture-of-Experts (MoE) architecture and a new "Lightning Attention" mechanism, enhancing long text processing capabilities and reducing computational resource requirements [2][3] - The training phase for the M1 model was completed in three weeks using 512 NVIDIA H800 GPUs, costing approximately $534,700 (around 3.84 million RMB) [3] Pricing Strategy - MiniMax employs a tiered pricing strategy for its API services, mirroring the pricing model of another model, Doubao, with specific rates based on input length [4][5] - The pricing for M1 is as follows: - $0.8 per million tokens for input lengths of 0-32k, $8 for output - $1.2 per million tokens for input lengths of 32k-128k, $16 for output - $2.4 per million tokens for input lengths of 128k-1M, $24 for output [3][4] Market Implications - The tiered pricing model is expected to facilitate broader adoption of multi-modal deep thinking models, as it reduces token consumption costs, which is crucial for AI agents executing tasks [5] - MiniMax's business model focuses on pure API services for B-end clients, contrasting with other leading companies that offer customized services [5]