![DeepSeek]()
Credit
Chinese AI lab DeepSeek has unveiled DeepSeek-R1, an open-source reasoning model that claims to perform on par with OpenAI’s o1 on key benchmarks. Available on Hugging Face under an MIT license, R1 can be used commercially without restrictions. DeepSeek asserts that R1 outperforms o1 on benchmarks like AIME, MATH-500, and SWE-bench Verified, which test reasoning, problem-solving, and programming capabilities.
As a reasoning model, R1 focuses on accuracy by fact-checking its outputs, though it takes longer to process results compared to nonreasoning models. The trade-off, DeepSeek claims, is improved reliability in areas like science, physics, and math. The full model features 671 billion parameters, enhancing its problem-solving abilities. Distilled versions ranging from 1.5 billion to 70 billion parameters have also been released, allowing smaller versions to run on laptops. The full model, requiring advanced hardware, is available via API at prices up to 95% lower than OpenAI’s o1.
However, R1 faces challenges as a Chinese-developed model, subject to government regulations enforcing “core socialist values.” It declines to respond to topics like Tiananmen Square or Taiwan’s autonomy, aligning with Chinese internet policies.
The release comes amid rising U.S.-China tensions over AI. The Biden administration has proposed stricter export restrictions on advanced AI technologies for Chinese firms. OpenAI recently urged the U.S. to bolster domestic AI development, singling out DeepSeek’s parent company as a competitor of concern.
DeepSeek joins Alibaba and Moonshot AI’s Kimi in claiming to rival OpenAI’s capabilities, with R1 marking a significant milestone in China’s rapid advancements in AI. Critics note that this trend highlights China’s strategy of quickly catching up to global leaders in AI innovation.