China’s AI is catching up: DeepSeek releases R1 model to challenge U.S. technology leadership

The Chinese AI research lab DeepSeek recently released its open-source reasoning model DeepSeek-R1, which has generated widespread industry attention. The model is referred to as a "reasoning model" and is said to perform on par with OpenAI's o1 in certain AI benchmarks. R1 has been released through the AI development platform Hugging Face under the MIT license, allowing users to commercialize it without restrictions.

DeepSeek claims that R1 outperforms o1 in several benchmarks, including the American Invitational Mathematics Examination (AIME), MATH-500, and SWE-bench Verified. AIME uses other models to assess reasoning ability, MATH-500 focuses on word problems, and SWE-bench Verified tests programming tasks.

The R1 model has advantages, but is limited by politics

As a reasoning model, R1 is said to have a unique self-verification capability, making it more reliable than traditional models in fields like physics, science, and mathematics. Although reasoning models generally require longer computation times (from seconds to minutes), their high accuracy gives them a great advantage in handling complex problems.

The technical report indicates that R1 contains 671 billion parameters, far exceeding many existing models. The number of parameters is usually proportional to a model's problem-solving capability, making R1 a large-scale model. However, DeepSeek also provides "slimmed-down" versions of R1 with parameters ranging from 150 million to 70 billion, with the smallest version even able to run on a laptop. The full version of R1 requires higher-performance hardware, but its cost through DeepSeek's API is only 5%-10% of OpenAI's o1.

However, R1's Chinese background also brings some limitations. Due to China's internet regulation, R1 must adhere to the "core socialist values" and thus refuses to respond to certain sensitive topics (such as the Tiananmen Square incident or Taiwan independence). This is a common feature of many Chinese AI systems, with other models like Moonshot's Kimi also subject to similar constraints.

China's AI "overtaking on the curve" and open-source supply

DeepSeek's R1 model highlights China's rapid progress in AI innovation and resource optimization. Facing the challenge of US restrictions on the export of advanced chips, Chinese AI developers have adopted technologies such as Mixture-of-Experts (MoE) to optimize the allocation of computing resources. DeepSeek has also previously launched an AI training cluster called Fire-Flyer 2, which combines around 10,000 Nvidia A100 chips and has lower costs and energy consumption than similar international systems.

DeepSeek's "slimmed-down" R1 models are another important breakthrough. These models not only can run on local hardware, but are also provided in open-source form, making AI technology more accessible and widespread globally. As George Mason University AI researcher Dean Ball pointed out, "The proliferation of these efficient reasoning models means they will continue to exist in a decentralized way, away from any top-down control mechanisms."

Export control: the US is wary of DeepSeek

Despite technical progress, China's AI industry still faces significant challenges. At the end of 2023, the US government further tightened export controls on China-related AI technologies and chips. In addition to the previous ban on advanced chips, the new rules may also limit Chinese companies' access to key semiconductor technologies needed to power advanced AI systems.

In response, OpenAI has warned in a policy document that the US government should strengthen support for domestic AI technology to prevent Chinese models from surpassing the US in capability. In an interview with The Information, OpenAI's policy vice president Chris Lehane even named DeepSeek's parent company High Flyer Capital Management as an "organization of particular concern".

The scale of China's AI companies is still far behind the US

Despite DeepSeek's impressive technical performance, the valuations of Chinese AI companies are still far behind their US counterparts. For example, OpenAI's valuation has reached $157 billion, while companies like DeepSeek and Moonshot are valued at only a fraction of that. Fierce market competition has led to price wars, squeezing the profit margins of AI model providers.

Beijing-based Zhipu AI is an example. The company is valued at around $3 billion, far behind OpenAI, and has therefore postponed its planned IPO in the second half of 2025. Although Zhipu recently demonstrated an image generation model similar to OpenAI's Sora, its commercial prospects remain unclear.

The future of China's AI: breaking through the constraints

China's AI industry is trying to break through the constraints of US restrictions and resource shortages. The rapid progress of DeepSeek's R1 model and other startups demonstrates China's immense potential in the AI field. As a former Beijing AI industry insider said, "We're dancing with shackles on, only able to focus on our strengths, and that's the only chance for survival and possible victory."

The future of China's AI will depend on the combination of its innovation capabilities and commercialization potential in the global competition. With the advancement of technology, companies like DeepSeek have proven China's ability to "overtake on the curve" in the AI field, but the global policy and market environment will continue to affect the outcome of this technology race.