New Major Study Suggests DeepSeek Still Lags Behind Top American Models and Presents Major Security Flaws

In a head-to-head matchup over the future of artificial intelligence (AI), the United States appears to be winning — for now. On September 30, the Center for AI Standards and Innovation at the National Institute for Standards and Technology (NIST) released a study suggesting that leading American AI models outperform DeepSeek, a top Chinese AI firm, on a range of industry-specific benchmarks designed to mimic common queries in programming, mathematics, science, and other fields.

But while the study found that American models remain superior at a range of engineering-related tasks, it also illustrated DeepSeek’s dramatic progress under intense American pressure, highlighting the insufficiency of current export controls on advanced semiconductors and related components.

DeepSeek’s Models Show Significant Improvement Over Previous Iterations

In evaluating DeepSeek’s models’ performance on a range of programming tasks, NIST found that the firm’s most advanced systems lagged behind their American competitors across a series of key cyber-related and software engineering benchmarks — particularly those related to cryptography and solving coding challenges. NIST also found that DeepSeek’s models suffered from significant security flaws, with testers able to easily compel their compliance with overtly malicious user queries, such as downloading and exfiltrating passwords or intentionally running malware.

However, DeepSeek’s models performed nearly at parity to American models in answering scientific questions and mathematical problems, both of which highlight DeepSeek’s overall approach in using reasoning models to solve complex user queries. Moreover, the report also highlighted that more recent DeepSeek models far outperform previous iterations, suggesting that the firm has managed to drastically improve its products even under intense pressure from U.S. export controls on advanced semiconductors used for training.

China’s AI Sector Continues To Rely on American Components

Despite its advances, DeepSeek, as with much of China’s AI sector, remains reliant on American components, particularly advanced semiconductors designed for both training and deploying models at scale. Shortly after the launch of its initial models in January, DeepSeek came under investigation by American officials for evading U.S. export controls to access large stores of Nvidia H100 chips acquired via shell companies in Southeast Asia. While the allegations remain unproven, Singaporean officials arrested several individuals accused of smuggling Nvidia chips to DeepSeek in February, while Malaysian authorities have investigated whether Chinese AI companies used data centers based in the country to train large language models. There are no restrictions on the sale of these chips to Malaysia.

This trend aligns with China’s overall AI development trajectory — models that demonstrate gradual improvement on specific tasks built primarily on American hardware. While the performance of Chinese open-source models has improved over time, these improvements have been primarily aimed toward accomplishing specific goals — such as running assembly lines — rather than developing systems that display generalized intelligence. Though this tendency has been driven by Beijing’s overarching desire to develop an advanced, automated manufacturing economy, it has also been partially driven by deficiencies within China’s overall AI ecosystem, which lacks the components required to run advanced, generalized models at scale.

Washington Should Step Up Efforts To Broaden and Enforce Export Controls

Even as it continues to lag behind American competitors, DeepSeek’s improvements offer a warning to Washington that China’s AI sector will continue to be a pacing competitor, particularly in practical applications such as aiding scientific discovery.

In response, the United States should broaden its export controls on key inputs while stepping up its enforcement measures. The Commerce Department should expand restrictions on the export of chip design software and critical minerals used in the production process, while Congress should consider legislation to screen outbound investment into China’s technology sector. Moreover, the Trump administration should rescind its hiring freeze to allow Commerce to increase its enforcement efforts, particularly following an unprecedented expansion of its export control infrastructure.

Jack Burnham is a research analyst in the China Program at the Foundation for Defense of Democracies (FDD). For more analysis from Jack and FDD, please subscribe HERE. Follow Jack on X @JackBurnham802. Follow FDD on X @FDD. FDD is a Washington, DC-based, nonpartisan research institute focusing on national security and foreign policy.