DeepSeek is the open-source Chinese AI lab that shook up the industry in early 2025 with its R1 reasoning model, delivering frontier-level performance at a fraction of the typical training/inference cost. Its current flagship generation is DeepSeek V4 (Pro and Flash variants), previewed in April 2026 and moving to official release with peak/off-peak API pricing in mid-2026. V4 Pro is a 1.6-trillion-parameter Mixture-of-Experts model (49B active parameters) with a 1-million-token context window and a new "Hybrid Attention Architecture" aimed at improving memory across long conversations; V4 Flash is a smaller, cheaper 284B-parameter variant. Both emphasize stronger agentic and coding capability plus large-context efficiency compared to the prior V3 series.