Monday, 30 June 2025
29.3 C
Singapore
32.2 C
Thailand
20.6 C
Indonesia
28.5 C
Philippines

DeepSeek claims its ‘reasoning model’ outperforms OpenAI’s o1 on key benchmarks

DeepSeek’s R1 claims to outperform OpenAI’s o1 in reasoning tasks, but regulatory and geopolitical issues shape its limitations and potential impact.

Chinese AI lab DeepSeek has unveiled its reasoning model, DeepSeek-R1, which it says rivals OpenAI’s o1 on several key AI benchmarks. The model, now available on the AI development platform Hugging Face under an MIT license, is open for commercial use without restrictions.

DeepSeek claims that R1 surpasses o1 performance on benchmarks such as AIME, MATH-500, and SWE-bench Verified. AIME evaluates models using other models, MATH-500 tests word problem-solving, and SWE-bench Verified assesses programming tasks.

How R1 works and what sets it apart

R1 is designed as a reasoning model, meaning it checks its work to avoid common pitfalls faced by typical AI systems. While this self-checking process takes slightly longer — often seconds to minutes more — it ensures higher reliability, especially in science, mathematics, and physics.

The model boasts an impressive 671 billion parameters, significantly enhancing its problem-solving capabilities. For comparison, models with more parameters are typically better at understanding and solving complex problems. Alongside the full version of R1, DeepSeek has also released smaller “distilled” versions, ranging from 1.5 billion to 70 billion parameters. The smallest versions are light enough to run on a standard laptop, while the full-scale R1 requires robust hardware.

For developers who need access to the full R1 but lack the necessary infrastructure, DeepSeek offers the model through its API at costs 90%-95% lower than those of OpenAI’s o1, making it an attractive option for many users.

Challenges and geopolitical implications

However, DeepSeek’s Chinese origins bring certain limitations. The model’s outputs must comply with regulations imposed by China’s internet watchdog, ensuring that its responses align with “core socialist values.” This means R1 avoids answering politically sensitive topics, such as Tiananmen Square or Taiwan’s autonomy. Many other Chinese AI models also avoid controversial discussions to remain in compliance with the government.

The launch of R1 coincides with rising tensions between the U.S. and China over AI technology. Recently, the Biden administration proposed stricter export rules, limiting China’s access to advanced AI chips and models. These rules would tighten existing restrictions on the tools needed to develop cutting-edge AI systems if implemented.

In a policy recommendation last week, OpenAI urged the U.S. government to prioritise American AI development to maintain its competitive edge. Chris Lehane, OpenAI’s VP of policy, identified DeepSeek’s parent company, High Flyer Capital Management, as a competitor to watch.

A growing trend in Chinese AI

DeepSeek is not alone in challenging U.S. dominance in AI. Other Chinese labs, such as Alibaba and Moonshot AI’s Kimi, have also developed models they claim rival OpenAI’s o1. DeepSeek, however, was the first to preview its reasoning model, R1, back in November.

Dean Ball, an AI researcher at George Mason University, noted that these developments suggest Chinese AI labs are becoming “fast followers.” He highlighted the accessibility of DeepSeek’s distilled models, which allow powerful reasoning capabilities to operate on local hardware.

With models like R1, Chinese AI firms continue to push boundaries despite regulatory challenges and geopolitical tensions.

Hot this week

Meta’s copyright victory raises questions about AI fair use

Meta won a copyright case over AI training, but the judge says future rulings could differ if better arguments are made.

Microsoft to announce major Xbox layoffs next week

Microsoft is preparing to cut more Xbox jobs next week as part of a larger restructure. Layoffs are also expected in its sales division.

Google launches Gemini AI for schools and students, raising questions about future of learning

Google launches Gemini AI in schools with safety tools and fact-checking, sparking debate on its impact on learning and student development.

AWS and DISG launch AI Springboard to help 300 Singapore enterprises adopt AI

AWS and DISG launch AI Springboard to help 300 Singapore enterprises scale AI adoption with funding, training, and technical support.

Nearly half of online registrations now flagged as attacks, says identity report

Nearly 50% of online registrations in 2024 were flagged as attacks, according to a new customer identity report by Auth0.

Android 16 to alert you if your phone connects to a fake cell tower

Android 16 will warn you if your phone connects to a fake tower, helping protect your calls, texts, and location from silent spying.

YouTube Create is finally coming to iOS devices

YouTube Create is coming to iOS, offering free mobile video editing tools as Google aims to catch up with CapCut and InShot.

OpenAI turns to Google’s AI chips in the shift from Microsoft and Nvidia

OpenAI begins renting Google's AI chips to run ChatGPT, shifting away from Microsoft and Nvidia to lower computing costs.

Google launches Gemini AI for schools and students, raising questions about future of learning

Google launches Gemini AI in schools with safety tools and fact-checking, sparking debate on its impact on learning and student development.

Related Articles

Popular Categories