DeepSeek
DeepSeek · founder Liang Wenfeng
Reach the frontier through efficiency, and publish the weights when you get there.
Why it exists
DeepSeek is the outlier at this table: no venture round, no consumer ambitions at the start, and a founder who made his money in quantitative trading. Liang Wenfeng's fund High-Flyer had bought thousands of GPUs for market prediction, and in July 2023 he spun the AI work out into its own lab — funded from trading profits rather than investors.
Export controls that were meant to hold Chinese labs back pushed this one into a different discipline. DeepSeek's papers are unusually specific about how they squeezed more out of less: attention that costs less memory, mixture-of-experts that activates a fraction of the model, training runs reported in single-digit millions of dollars when Western equivalents cost hundreds.
R1, released in January 2025 with open weights, is when the rest of the world noticed. A model that matched the frontier on reasoning and could be downloaded for free knocked roughly a trillion dollars off US tech stocks in a single day — and permanently changed the argument about what a frontier model has to cost.
The company
- Founded
- 17 Jul 2023
- Headquarters
- Hangzhou, China
- Founders
- Liang Wenfeng (founder of the quant fund High-Flyer)
- Structure
- Private lab funded by the quant fund High-Flyer
- Stated mission
- Reach the frontier through efficiency, and publish the weights when you get there.
Go to the source
Only values the vendor published are plotted. Where a model has no stated context window, there is no point on the line — nothing is estimated.
Only values the vendor published are plotted. Where a model has no stated context window, there is no point on the line — nothing is estimated.