Over the weekend, DeepSeek, a Chinese AI company based in Hangzhou, made headlines by releasing a new AI chat app. Their new model, called R1, is described as a “reasoning” AI, similar to OpenAI’s o1. This launch sent shockwaves through the American tech scene, and DeepSeek shot to the top of Apple’s App Store.
DeepSeek has already gained attention in the U.S. market with its offerings, like DeepSeek-V3, which resembles GPT-4 in capabilities. These models are designed to quickly process natural language prompts and deliver answers. Following the app’s release, both NVIDIA and Microsoft saw declines in their stock prices on Monday, showing a dip in investor confidence in American AI companies. This situation raises questions about whether U.S. restrictions on chip access for Chinese firms stifle or foster competition.
For those in tech, DeepSeek represents a new avenue for coding and enhancing productivity. The R1 model stands out for its ability to explain its reasoning process and is part of an open-source ecosystem available on GitHub.
DeepSeek-V3 and R1 are built to “reason through” their output, which leads to more accurate results, especially in math and coding challenges. DeepSeek claims their V3 model outperformed GPT-4o on key benchmarks like the MMLU and HumanEval tests. Remarkably, training one of their models cost $5.6 million, which is much lower than typical expenditures in Silicon Valley.
Users can access DeepSeek’s models via the App Store or through web browsers, with the R1 model providing detailed, conversational explanations for its answers. Though the website noted potential service disruptions, the chatbot remained functional as of Monday morning.
DeepSeek also offers an API that works with the OpenAI SDK, allowing developers to integrate its capabilities into their applications.
According to Gartner analyst Arun Chandrasekaran, the introduction of DeepSeek’s V3 and R1 models could lead to a broad array of applications built around the R1 model with support from global cloud providers. Success for DeepSeek hinges on continual innovation and establishing a developer ecosystem while navigating cultural challenges.
With V3 trained on 2,048 NVIDIA H800 GPUs, DeepSeek highlights an interesting point: U.S. companies can’t sell high-performance AI training chips to Chinese firms under current export regulations. Ivan Feinseth of Tigress Financial remarked that DeepSeek’s potential, coupled with its low costs, challenges the massive investments made in the U.S. AI sector.
DeepSeek is also gaining ground as an open-source, research-driven initiative, while competitors like OpenAI are focusing more on commercial goals. Venture capitalist Marc Andreessen expressed his admiration for DeepSeek R1, calling it an impressive breakthrough and a meaningful contribution to the world.
Finally, on Monday, DeepSeek announced the Janus-Pro family of multimodal models, which can process and generate images, adding another layer to their innovative offerings.
