DeepSeek: deepseek-v3-0324

A state-of-the-art 671B MoE model (37B active) featuring Multi-head Latent Attention (MLA) and auxiliary-loss-free load balancing. It provides a hybrid "Think/Non-think" mode, delivering SOTA performance in coding, math, and agentic tasks while remaining the most cost-efficient flagship in the 2026 market.