How did DeepSeek R1 shock Silicon Valley with open-source reasoning?

The answer

On January 20, 2025, Chinese AI lab DeepSeek released R1, an open-source reasoning model under the MIT license that matched OpenAI o1's performance across math, code, and reasoning benchmarks. R1 scored 79.8% on AIME and 97.3% on MATH-500, rivaling o1 at a fraction of the cost. Built on a 671-billion-parameter Mixture-of-Experts architecture with only 37 billion parameters active per token, its most remarkable feature was DeepSeek-R1-Zero - a variant trained with pure reinforcement learning that spontaneously developed chain-of-thought reasoning without any supervised fine-tuning. DeepSeek also released distilled versions from 1.5B to 70B parameters, democratizing advanced reasoning capabilities for the broader AI community.