In a development that has sent shockwaves through the tech industry, Chinese startup, DeepSeek, has unveiled its latest artificial intelligence (AI) model, DeepSeek-R1.
DeepSeek-R1 is designed to handle a variety of text-based tasks in both English and Chinese, including creative writing, general question answering, editing, and summarization. According to DeepSeek, the model is particularly strong in reasoning-intensive tasks that require clear, well-defined solutions, such as generating and debugging code, performing mathematical computations and explaining complex scientific concepts.
A New Benchmark in AI Performance
In a 2025 performance evaluation, published by Statista, DeepSeek-R1 demonstrated impressive results, performing on par with OpenAI's OpenAI-01-1217. This has drawn significant attention as DeepSeek's model was developed with a substantially lower budget.
DeepSeek-R1 is now available on the AI development platform, Hugging Face, under an MIT license, allowing unrestricted commercial use. DeepSeek claims that R1 outperforms OpenAI’s o1 as per the following benchmarks: AIME, MATH-500, and SWE-bench Verified. AIME evaluates AI performance using other models, MATH-500 comprises a set of word problems, and SWE-bench Verified assesses programming capabilities.
As a reasoning model, R1 is designed to fact-check itself, reducing the risk of common errors that often plague AI systems. While reasoning models generally take longer—sometimes several seconds to minutes—to generate solutions compared to non-reasoning models, they offer greater reliability in fields such as physics, science, and mathematics.
DeepSeek’s technical report revealed that R1 features 671 billion parameters, a measure indicative of a model’s problem-solving capacity. Generally, AI models with a higher parameter count deliver superior performance.
Recognizing the need for scalability, DeepSeek has also introduced “distilled” versions of R1, with parameter sizes ranging from 1.5 billion to 70 billion. The smallest version is capable of running on a standard laptop, while the full-scale R1 demands more robust hardware. Nevertheless, DeepSeek provides access to R1 through its API at costs 90%-95% lower than OpenAI’s o1.
Hugging Face CEO, Clem Delangue, revealed in a post on X that developers on the platform have already created over 500 derivative models based on R1. These variations have amassed 2.5 million downloads, five times the number of downloads for the official R1 release, highlighting the model’s rapid adoption and growing influence in the AI community.
Also Read: GenAI Sparks Telco-Hyperscaler Collaboration
Insights into the Asian Landscape
DeepSeek's breakthrough is part of a broader trend whereby Asian companies are making significant strides in AI development. A 2024 report by the Asian Development Bank highlighted several key insights:
- Emerging Regional AI Hubs: Beyond China, countries like South Korea, Japan, and Singapore are becoming AI powerhouses. Each is leveraging its unique strengths in electronics, robotics, and financial technology to carve out niches in the global AI ecosystem.
- Sector-Specific Innovations: Asian companies are increasingly focusing on domain-specific AI applications. For instance, South Korea's Lunit has become a global leader in AI-powered medical imaging, with its Lunit INSIGHT CXR demonstrating a 97% accuracy rate in detecting early-stage lung cancer.
- Ethical AI Frameworks: Asian countries are taking the lead in developing comprehensive ethical guidelines for AI. Singapore's Model AI Governance Framework, launched in 2020, has since been adopted by over 15 countries in the region, setting a standard for responsible AI development and deployment.
China's Growing AI Prowess
The success of DeepSeek-R1 is emblematic of China's rapidly advancing AI capabilities. China leads the world in generative AI (GenAI) patent filings, surpassing the U.S., South Korea, Japan, and India, according to a World Intellectual Property Organization (WIPO) Patent Landscape Report on Generative AI.
The study documented 54,000 GenAI inventions between 2014-2023, with over 25% emerging in the last year alone. China accounts for more than 38,000 of these inventions—six times more than the U.S. India, ranked fifth, has the fastest growth rate at 56% annually.
Industry News: China to Build 10,000 5G-Powered Factories to Transform Industries
DeepSeek Tops App Store, Triggers Historic NVIDIA Stock Plunge
DeepSeek's chatbot, powered by R1, has risen to the top spot on the Apple App Store, overtaking ChatGPT. This success challenges the notion that only the largest, most expensive models can achieve state-of-the-art performance, potentially leading to a shift in focus towards more efficient training methods and architectures.
The open-source nature of R1 is likely to spur collaborative improvements and adaptations, potentially accelerating the overall pace of AI innovation. Furthermore, DeepSeek may intensify the ongoing "AI race" between nations, potentially leading to increased government investment and regulation in the field.
The announcement of R1 also initially caused significant market reactions, including a historic drop in NVIDIA's stock price. This occurred after DeepSeek revealed it developed R1 using just USD 5.6 million worth of NVIDIA GPUs, challenging the chip giant's pricing model. However, recent market analyses suggest a stabilization as investors reassess the long-term implications of DeepSeek's innovations and NVIDIA's role in the AI hardware ecosystem.
Investigations into Hardware Procurement
Reports indicate that U.S. authorities have launched an investigation into DeepSeek over suspicions of circumventing export restrictions on advanced NVIDIA GPUs. The probe centers on whether DeepSeek acquired restricted AI chips through intermediaries in Singapore, effectively bypassing U.S. export controls.
The investigation comes amid growing concerns about DeepSeek's AI model, R1, which has demonstrated capabilities rivaling those of leading models from OpenAI and Google. Adding to the suspicion, Singapore's share of NVIDIA's revenue has more than doubled from 9% to 22% over the past two years.
While DeepSeek has not disclosed the specific hardware used to train R1, the company's previous model, V3, was reportedly trained using a limited number of H800 GPUs. This achievement was notably more efficient than similar efforts by competitors like Meta, raising questions about the resources behind R1's development.
U.S. officials, including representatives from the White House and FBI, are examining whether DeepSeek obtained restricted AI GPUs through Singaporean third-party firms. NVIDIA maintains that it complies with all legal requirements, and no official confirmation of law violations has been made public.
The Biden administration tightened restrictions on advanced GPU exports to China in 2023, however, Singapore was not included in the list of restricted countries. This omission has led some to believe it created a loophole for Chinese entities to access high-end NVIDIA GPUs.
In response to these concerns, U.S. representatives, John Moolenaar and Raja Krishnamoorthi, have called for stricter licensing measures unless Singapore enhances its oversight on shipments. NVIDIA has clarified that while Singapore accounts for a significant portion of its revenue, most transactions involve shipments to other destinations, not necessarily China.
Network News: Indosat Ooredoo Hutchison Partners with NVIDIA for Sovereign AI Infrastructure
Competition Heats Up
South Korean tech company, Kakao, known for its dominance in online banking, ride-hailing, and messaging services, has recently joined forces with OpenAI to integrate ChatGPT into its AI offerings. The partnership comes amid heightened competition, particularly with the emergence of DeepSeek.
During a private discussion with South Korean AI developers, OpenAI CEO Sam Altman acknowledged that OpenAI has yet to formulate a clear strategy to counter the rise of open-source models like DeepSeek. Despite this, he admitted that OpenAI is closely monitoring the situation and considering ways to adapt.
Experts suggest that South Korea, with its highly engaged and tech-savvy user base, could serve as a testing ground for OpenAI before broader global expansion. Meanwhile, Altman has been meeting with major South Korean semiconductor manufacturers, including Samsung and SK hynix, to explore collaborations on AI memory chips, particularly high-bandwidth memory (HBM), which is crucial for AI applications.
DeepSeek’s rapid advancement has also sparked allegations that it may have replicated elements of leading U.S. AI technologies. OpenAI has warned that Chinese firms are actively working to emulate its models using a technique called distillation, where smaller AI systems learn by mimicking the behavior of larger, more advanced models. At the same time, OpenAI itself faces ongoing legal challenges over alleged copyright violations in training its generative AI models.
DeepSeek-R1’s Alarming Vulnerabilities
Using algorithmic jailbreaking techniques, Cisco tested DeepSeek-R1 against 50 random prompts from the HarmBench dataset, which evaluates AI models on their ability to reject harmful behaviors like cybercrime, misinformation, and illegal activities.
DeepSeek-R1 reportedly exhibited a 100% attack success rate, meaning it failed to block a single harmful prompt. This stands in stark contrast to other leading AI models, which demonstrated at least some resistance to attacks.
According to Cisco, other frontier models such as o1 successfully blocked a majority of adversarial attacks due to built-in guardrails. One potential reason for DeepSeek’s poor performance is its significantly lower development budget.
Despite its vulnerabilities to adversarial attacks, DeepSeek-R1 enforces strict content restrictions on politically sensitive topics, particularly concerning China. When tested on issues like the treatment of Uyghurs or the 1989 Tiananmen Square Massacre, DeepSeek refused to respond, stating, "Sorry, that's beyond my current scope. Let’s talk about something else."
Global Scrutiny Over Emerging AI Technologies
Australia has banned the use of DeepSeek across all government systems following a risk assessment that flagged potential national security threats. Home Affairs Minister Tony Burke announced the immediate removal of all DeepSeek products from government networks, emphasizing that the decision was based on security concerns rather than the company’s Chinese origins. While the restriction applies only to government devices, Burke urged Australian citizens to remain cautious about their digital security. This move is reminiscent of Australia’s previous actions against Chinese tech firms, such as the 2018 exclusion of Huawei and ZTE from the country’s 5G infrastructure over security risks.
The U.S. is also considering similar measures. Republican Senator Josh Hawley has introduced a bill that would prohibit individuals in the U.S. from using DeepSeek or engaging with any AI technology developed in China, with potential penalties including fines of up to USD 1 million or prison sentences of up to 20 years. While the U.S. has not yet imposed an official ban on government devices, congressional offices have reportedly been advised against using the platform. Agencies like NASA and the Pentagon’s Defense Information Systems Agency have already restricted access, while the U.S. Navy has imposed a blanket ban on its personnel, citing security and ethical concerns. Texas has also become the first U.S. state to prohibit DeepSeek and other Chinese software on government devices.
Beyond the U.S. and Australia, other nations have taken action. Italy’s Data Protection Authority issued a nationwide ban, citing potential data breaches and requiring the company to prove compliance with the General Data Protection Regulation (GDPR).
Data protection authorities in Belgium, France, and Ireland are investigating DeepSeek’s data handling practices, while South Korea and India have indicated that regulatory action may follow due to privacy and security concerns
Taiwan’s government agencies have also prohibited the AI platform, warning that it poses risks to national information security.
In the Philippines, the Armed Forces (AFP) has explicitly prohibited its personnel from uploading sensitive information to AI-powered platforms, including DeepSeek, to prevent potential national security breaches. During a recent congressional hearing, lawmakers questioned the military’s stance on AI technologies, with officials emphasizing the need for caution. Brigadier General Constancio Espina II noted that while DeepSeek is not explicitly banned, overlooking its risks could have serious security implications.
The growing global response to DeepSeek highlights rising concerns over AI security. As AI technology advances, addressing these security challenges will be crucial to ensure safe and reliable integration into various industries.
Read More:
The Time is Now: Artificial Intelligence Spending Projected to Rise in Asia Pacific
Artificial Intelligence in Asia: Can AI Threaten Jobs in ASEAN Countries?
Great feature!
