DeepSeek Unveils New AI Model in China—Key Details to Consider
Overview of DeepSeek’s Latest AI Model Update
DeepSeek, a burgeoning AI startup from China, has recently rolled out an upgrade to its large language model. The latest version, named DeepSeek-V3-0324, claims “significant improvements” compared to its earlier model, intensifying its competition with major AI players like OpenAI. This update follows a previous release that notably influenced global technology sectors.
DeepSeek announced that its upgraded model shows “significant enhancements” in benchmarking metrics.
Source: AFP via Getty Images
Key Highlights of the Update
On Tuesday, DeepSeek unveiled its upgraded model on the Hugging Face AI development platform. The new version boasts improvements in reasoning and coding functionalities compared to the original V3 model.
The original DeepSeek V3 was launched on December 26, 2024, and has been noted for being three times faster than its predecessor, placing it among the top 10 on Chatbot Arena, an evaluation platform managed by the University of California, Berkeley.
The latest version includes notable advancements in various benchmark tests, enhancements in front-end web development, proficiency in Chinese writing, and improved features like “interactive rewriting” and “enhanced report analysis.”
Market Reactions
Observers are closely watching how DeepSeek’s recent advancements might impact U.S. tech stock prices. Following the model’s announcement, shares of Nvidia saw a decline of 1.2% and other prominent stocks like Broadcom and Tesla also faced slight dips. Conversely, companies like Apple, Meta, and Microsoft experienced minor gains.
Understanding DeepSeek
Founded in 2023 by Liang Wenfeng, DeepSeek emerged from an ambition to contribute to the AI landscape in China. Liang notably acquired a substantial number of Nvidia graphics processors for an undisclosed AI project in 2021, just ahead of restrictions on chip sales to China by the U.S. government. He advocates for China’s shift towards becoming a leader in AI technology rather than just following existing innovators. DeepSeek claims its models are both efficient and cost-effective, asserting that developing one of its latest models cost around $5.6 million compared to competitors like OpenAI, who may spend between $100 million and $1 billion.
Regulatory Attention on DeepSeek
Recent developments in U.S. policy have flagged potential concerns regarding DeepSeek’s operations. A bipartisan bill has been introduced to ban the use of DeepSeek on federal devices due to its alleged ties to a sanctioned Chinese telecommunications entity. Although uncertainty surrounds legislative support for the bill, certain U.S. government agencies, including the Navy and Defense Department, have already limited access to DeepSeek’s technology due to security considerations. Congressional warnings have also surfaced cautioning against the use of DeepSeek, attributing its products to possible malicious software exploits.
Background on DeepSeek’s Technologies
Earlier this year, DeepSeek had launched its advanced reasoning model R1, which positioned the company in direct competition with U.S. AI giants, achieving results comparable to OpenAI’s o1 model in multiple benchmarks. The R1 model is available as open-source software, which has gained immense popularity, even overtaking OpenAI’s ChatGPT in download rankings on platforms like the iPhone. The introduction of DeepSeek’s R1 model evoked market shifts, leading to substantial losses in shares of U.S. tech companies such as Nvidia, which experienced a staggering 17% drop in one day.
