DeepSeek claims its ‘reasoning’ model beats OpenAI’s o1 on certain benchmarks

Chinese AI lab DeepSeek has launched an open-source version of its reasoning model, DeepSeek-R1, which is now available on the Hugging Face platform under an MIT license. The model, which boasts 671 billion parameters, outperforms OpenAI’s o1 on several AI benchmarks, including AIME, MATH-500, and SWE-bench Verified. The reasoning model is designed to be more reliable in fields like physics, science, and math, despite taking longer to solve problems. DeepSeek also offers smaller versions of R1, ranging from 1.5 to 70 billion parameters, suitable for use on less powerful hardware.

R1’s open release marks a significant development, as it is available commercially without restrictions and comes at a fraction of the cost of OpenAI’s o1. However, the model has limitations due to Chinese internet regulations, leading it to avoid sensitive topics like Tiananmen Square and Taiwan. Despite these constraints, R1 has attracted significant attention, with developers creating over 500 derivative models. R1’s release comes amid growing concerns in the U.S. about Chinese AI development, particularly in response to new export restrictions on AI technologies.

This launch also positions DeepSeek among a growing list of Chinese labs, including Alibaba and Kimi, that have developed models capable of rivaling OpenAI’s systems. Experts suggest that Chinese AI labs will continue to advance rapidly, producing highly capable AI models that can run on local hardware, potentially bypassing regulatory controls and reshaping the global AI landscape.

Source: Techcrunch