2024年12月DeepSeek AI (中国Hangzhou・2023年High-Flyer Quantitative量化基金子会社設立・累計運用資金$50B+/年・中国AI startup leader・量化金融資金潤沢AI研究投資) 発表DeepSeek-V3・Industry-leading MoE scale Open weights LLM・671B total parameters (256 experts × 2.62B per expert) + 37B active params per token (Sparse activation・1/18 ratio・Industry-leading sparsity efficiency) + 128K context length + DeepSeek License (Permissive license・Commercial use可) + Multi-Token Prediction MTP + Auxiliary-loss-free load balancing + FP8 training・Industry-leading MoE scale + GPT-4o competitive performance + Cost $5.6M training (1/10 GPT-4 cost industry-shocking)。