DeepSeek V3 and R1 — MIT-Licensed Frontier AI, the RL Reasoning Breakthrough, and the DeepSeek Shock

DeepSeek V3 (December 2024) and R1 (January 2025) shook the AI industry: frontier-class performance, MIT license, $5.57M training cost disclosure, and an RL-only reasoning methodology that rivaled OpenAI o1. We review the architecture, benchmarks, the NVIDIA stock crash, and what it means for the open-weight AI landscape.