MiniMax M3.1 Flash beats DeepSeek with sub-100K H100-equivalent compute

MiniMax released M3.1 Flash with 428B total and 23B active parameters, priced at $0.15/$0.60 per 1M input/output tokens. It leads Automation Bench, Banker ToolBench and LOCA-Bench-256K despite Chinese labs having under 100,000 H100-equivalent GPUs.

• 428B total and 23B active parameters at $0.15/$0.60 per 1M tokens
• Tops Automation Bench, Banker ToolBench and LOCA-Bench-256K
• Per-token compute cut 20x; prefill 9x and decode 15x faster
• MiniMax and DeepSeek run on under 100,000 H100-equivalent GPUs