DeepSeek Ups Ante (Again) in Duel with OpenAI, Anthropic

By Jon Swartz

DeepSeek has raised the bar yet again in the artificial intelligence (AI) race.

The Chinese startup this week released a major upgrade to its V3 large language model (LLM) in its escalating jousting with American rivals OpenAI and Anthropic. The new model, DeepSeek-V3-0324, has 685 billion parameters, compared with the original V3 model’s 671 billion, offering “significant improvements” in reasoning and coding, the company said. The new model was made available through the Hugging Face platform.

The latest update, DeepSeek claims, boosts performance across several benchmark tests for the language model, as well as upgrades to front-end web development, Chinese writing proficiency and Chinese search capabilities such as “enhanced report analysis.”

DeepSeek’s latest version of V3, which roiled markets and heightened blood pressure across Silicon Valley when it was originally released late last year, is bound to again rattle U.S. tech companies that are investing billions of dollars in advanced chips and large data centers used to train AI models. Amazon.com Inc., Meta Platforms Inc., and Microsoft Corp. plan to spend $371 billion on data centers and computing resources this year, up 44% from a year ago.

When the Chinese company unveiled V3, it boasted the model was trained with less than $6 million worth of computing power from 2,000 NVIDIA’s H800 chips to achieve a level of performance on par with the most advanced models from OpenAI and Meta.

“If U.S. tech giants and AI companies continue on their current trajectory, they will be outpaced by their smaller, more agile counterparts who open-source their models and continue reducing costs and compute needs,” Tory Green, CEO of GPU network io.net, warns. “Despite incredible, government-backed efforts to protect their monopolies, there is likely little U.S. tech giants can now do to stem the tide of open-source AI development.”

So far, reviews of DeepSeek’s V3 have been generally positive among those in the tech industry, although a number of countries (Italy, Ireland, South Korea, Australia) and federal agencies have banned the use of the startup’s model out of concerns over its ties to the Chinese government.

Earlier this week during a visit to China, Apple Inc. CEO Tim Cook reportedly described DeepSeek’s AI models as “excellent” in an interview. When asked during an earnings call with analysts in January whether DeepSeek posed a risk to Apple’s future revenue, Cook said, “In general, I think innovation that drives efficiency is a good thing. And that’s what you see in that model.”

American technologists, in general, are increasingly worried about AI models coming out of mainland China. Since DeepSeek debuted V3 in late 2024, a conga line of companies have ramped up their AI plans: Baidu Inc. has introduced Ernie X1 to compete with DeepSeek’s R1 model; Alibaba Group Holding announced its own AI agents and reasoning model upgrade; Tencent Holdings unfurled an AI blueprint to rival R1; and Meituan, the world’s biggest meal-delivery service, said it was investing billions of dollars on AI.

“We’re witnessing a fundamental shift in the AI landscape. DeepSeek’s rapid development cycle — from launching last year to releasing multiple competitive models in quick succession — demonstrates China’s serious commitment to becoming a contributor rather than just a consumer in the global AI ecosystem,” David Bader, director of the Institute for Data Science at New Jersey Institute of Technology, said in an email.

“While regulatory concerns exist, with U.S. government restrictions already targeting DeepSeek’s products, the technical achievements can’t be ignored,” Bader added. “This isn’t just about national competition; it’s about how quickly AI capabilities are advancing globally and how different regulatory approaches might influence innovation trajectories in the years ahead.”

https://techstrong.ai/ai-at-the-edge/deepseek-ups-ante-again-in-duel-with-openai-anthropic/