This week, DeepSeek announced an upgrade to its large language model, which the China-based startup described as containing "significant improvements" compared to its predecessor.
Screenshot_20250325-090622.png
This development appears to intensify the competition with OpenAI and other U.S.
AI companies, following a previous release that unsettled global technology stocks. DeepSeek has rolled out its upgraded model, DeepSeek-V3-0324, on the Hugging Face AI development platform. This new iteration highlights enhanced reasoning and coding capabilities over its former V3 model. Originally launched on December 26, 2024, DeepSeek's V3 model was touted as being three times faster than the V2 version and secured a spot among the top 10 on Chatbot Arena, a performance evaluation platform at the University of California, Berkeley. The latest upgrade boasts notable improvements in various benchmark tests for the language model, alongside enhancements in front-end web development, proficiency in writing in Chinese—with new features such as "interactive rewriting"—and Chinese search capabilities like "enhanced report analysis," according to the company. What to Watch For: Observers are interested in how this latest model might influence U.S. tech stocks; as of 9:35 a.m. EDT, Nvidia shares fell by 1.2%, with other companies like Broadcom (down 0.6%) and Tesla (down 0.3%) also seeing slight declines. In contrast, Apple (up 0.3%), Meta (up 0.7%), and Microsoft (up 0.3%) experienced gains on Tuesday morning. About DeepSeek: Founded by entrepreneur Liang Wenfeng in 2023, DeepSeek began amassing thousands of Nvidia graphics processors for an unspecified AI project in 2021, prior to the Biden administration's restrictions on chip sales to China. Liang advocates for China to evolve into a significant contributor to the AI sector instead of merely following other nations. The company asserts that its products are more efficient and lower in development costs compared to offerings from OpenAI and Meta, estimating that the training cost for one of its latest models was around $5.6 million, significantly less than the $100 million to $1 billion suggested by an AI executive for developing comparable models. Will DeepSeek Be Banned in the U.S.?: A bipartisan bill aimed at prohibiting DeepSeek from federal devices was introduced in February following a report linking the company's chatbot to a Chinese state-run telecommunications firm that is banned in the U.S.
It is unclear how much support the bill will receive from other lawmakers. Additionally, the Trump administration is considering a ban due to national security concerns, according to sources familiar with the situation.
Several U.S. agencies, including the Navy, the Defense Department, the Commerce Department, and NASA, have restricted access to DeepSeek, with warnings issued that "threat actors are already exploiting DeepSeek to deliver malware and compromise devices." Background Information: In January, DeepSeek unveiled its advanced reasoning model, R1, which positioned it in direct competition with U.S. tech and AI leaders like OpenAI and Meta.
The R1 model performed comparably to OpenAI's o1 across multiple benchmarks and even surpassed it in the MATH-500 test. As an open-source model, R1 became widely accessible, surging to the top of the iPhone download charts, surpassing OpenAI's ChatGPT. The release of R1 resulted in a downturn for various U.S. tech stocks, with Nvidia experiencing a 17% drop on January 27, erasing $589 billion from its market capitalization—the largest single-day loss for any public company.
OpenAI CEO Sam Altman acknowledged the significant capabilities of DeepSeek's products, particularly praising the R1 model for its impressive performance relative to its cost.
Google