ÓÐÁϺÐ×ÓAPP

In January 2025, the introduction of a series of new artificial intelligence (AI) models by the under-the-radar Chinese developer DeepSeek upended the global AI market. Notably, DeepSeek claims to have matched the performance of the most advanced frontier models at a fraction of their development costs. Furthermore, the rapid launch of several models targeting specific use cases—general knowledge, reasoning, and image design—suggests this may be a new paradigm in AI development.

DeepSeek’s reported innovations raise important questions for policymakers and technologists alike, including:

1. Geopolitical Balance of Powers in AI: Most believed that the United States had a modest but sustainable advantage in AI development, enabled partially by constraints in hardware.
2. Economics of AI: Many AI leaders championed a "bigger-is-better" mantra regarding AI models and were creating a multibillion-dollar ecosystem to support this.
3. Business Model for AI: DeepSeek’s models are developed as open-source technology, providing potential cost benefits over proprietary models.

ÓÐÁϺÐ×ÓAPP’s A Technical Guide to DeepSeek assesses DeepSeek’s claimed innovations and explores their implications for AI’s future. Produced by ÓÐÁϺÐ×ÓAPP’s AI engineering research team, this new guide provides an in-depth technical assessment of DeepSeek’s large language models, including DeepSeek-R1, using published whitepapers and other resources. The guide explores and demystifies the many novel techniques DeepSeek used to optimize its end-to-end training and inference processes and explains how these innovations could be applied elsewhere.