Over the weekend, Chinese AI company DeepSeek launched an AI chat app featuring a “reasoning” AI model on par with OpenAI’s o1, causing a stir in the U.S. tech industry as it surged to the top of Apple’s App Store.
Based in Hangzhou, China, DeepSeek specializes in generative AI models and AI integration. Its first products to make a significant impact in the U.S. market include the GPT-4-like DeepSeek-V3 and R1, an advanced “reasoning model.” Much like ChatGPT, DeepSeek-V3 and R1 swiftly respond to natural-language prompts.
NVIDIA and Microsoft saw a drop in stock prices on Monday after the highly anticipated launch, reflecting a sudden dip in investor confidence in U.S. AI companies. DeepSeek’s rise has sparked debate on whether U.S. restrictions on Chinese access to AI chips are hindering or stimulating competition.
For tech professionals, DeepSeek offers an alternative for coding or boosting daily task efficiency. Alongside the R1 model’s ability to explain its reasoning, it’s based on an open-source family of models available on GitHub.
What sets DeepSeek apart?
Like OpenAI’s o1 (formerly Strawberry), the App’s reasoning model slows its prediction process to “reason through” its work, enhancing answer accuracy. Reasoning models, in particular, have excelled in benchmarks for math and coding.
DeepSeek claims that DeepSeek-V3 outperformed GPT-4o in the MMLU and HumanEval tests, which evaluate AI responses across a variety of tasks.
The company revealed that one of its models cost $5.6 million to train—a fraction of the typical investment in similar projects in Silicon Valley.
Both DeepSeek-V3 and R1 can be accessed via the App Store or on the web. Users visiting the DeepSeek site can select the R1 model for more detailed responses to complex queries. When chosen, the R1 model provides in-depth, conversational explanations of how it arrived at its conclusions.
As of Monday morning, the DeepSeek chat site displayed a warning of potential service disruption, though the chatbot was functioning as usual.
DeepSeek also offers an API that operates through the OpenAI SDK or compatible software.
What does DeepSeek’s launch mean for the AI industry?
“We can fully expect an ecosystem of applications to be built on R1, with several global cloud providers offering its models as a consumable API,” said Arun Chandrasekaran, Gartner Distinguished VP Analyst, in an email to TechRepublic. “DeepSeek’s long-term success will depend on its ability to continue innovating, build a developer ecosystem, and overcome cultural barriers given its country of origin.”
Chandrasekaran highlighted DeepSeek’s cost efficiency, strong benchmark results, and open weights as key differentiators.
DeepSeek-V3 was trained on 2,048 NVIDIA H800 GPUs. U.S. manufacturers are prohibited from selling high-performance AI training chips to Chinese companies under export rules set by the Biden administration.
“The potential power and low-cost development of DeepSeek raises questions about the hundreds of billions of dollars committed to U.S. companies,” noted Ivan Feinseth, a market analyst at Tigress Financial, in a note to clients obtained by ABC News.
It further sets itself apart by being an open-source, research-driven project, while OpenAI increasingly focuses on commercial ventures.
“DeepSeek R1 is one of the most impressive breakthroughs I’ve ever seen—and as open-source, it’s a profound gift to the world,” posted venture capitalist Marc Andreessen on X (formerly Twitter) on Friday.
Gartner forecasts the global AI semiconductor industry will reach $114 billion in 2025, and predicts the power needed to run newly added AI servers in data centers will hit 500 terawatt-hours by 2027.
DeepSeek introduces multimodal models
On Monday, DeepSeek unveiled another surprise: the Janus-Pro family of multimodal models, which can analyze and generate images.
TechRepublic
