US EditionSecrets of DeepSeek AI Model Revealed in Landmark Paper
DeepSeek-R1 used pure reinforcement learning to develop reasoning skills, costing just $294,000 to train, significantly less than rival models, the team said.
- Researchers at Chinese firm DeepSeek published a peer-reviewed paper in Nature on September 17, 2025, revealing details about their AI model R1.
- DeepSeek developed R1 by augmenting a base large language model with pure reinforcement learning, avoiding training on outputs from rival models like OpenAI's.
- The model was trained mainly on Nvidia’s H800 chips before US export controls banned such sales to China in 2023 and cost $294,000 to train, much less than competitors' tens of millions.
- R1 excels in reasoning tasks like mathematics and coding, has been downloaded 10.9 million times, and influenced reinforcement learning practices among AI researchers, as noted by experts like Huan Sun and Lewis Tunstall.
- The publication sets a precedent for transparency in AI development, suggesting that public sharing of training processes is key to evaluating risks and encouraging broader adoption of such methods.
17 Articles
17 Articles
DeepSeek reveals its popular AI R1 model cost just $294,000 to train, challenging US AI giants on price and chips - Tech Startups
Chinese AI startup DeepSeek has reemerged with a figure that could reignite one of the industry’s hottest debates: training costs. The company says its reasoning-focused R1 model cost just $294,000 to train — a number that undercuts U.S. rivals by […] The post DeepSeek reveals its popular AI R1 model cost just $294,000 to train, challenging US AI giants on price and chips first appeared on Tech Startups.
A small and unknown Chinese company, DeepSeek, revolutionized the generational artificial intelligence industry (AI) in January of this year. Its R1 model worked as well or better than the latest version of ChatGPT, but it was free and open source. It had been developed in China despite the embargo on chip exports and, as announced (and later questioned), with far fewer resources than the competition. DeepSeek’s team publishes today in Nature ma…
Coverage Details
Bias Distribution
- 50% of the sources lean Left
Factuality
To view factuality data please Upgrade to Premium
