DeepSeek: Everything you need to know about the AI chatbot app

DeepSeek has become a viral sensation.

Chinese artificial intelligence (AI) laboratory DeepSeek has rapidly gained mainstream recognition this week as its chatbot application soared to the summit of both the Apple App Store and Google Play Store charts. DeepSeek’s AI models, developed through compute-efficient techniques, have prompted inquiries from Wall Street analysts and technologists alike regarding the United States’ capacity to sustain its dominance in the AI race and the enduring demand for AI chips.

DeepSeek’s Origins in Trading

DeepSeek’s financial backing comes from High-Flyer Capital Management, a Chinese quantitative hedge fund utilizing AI to guide its trading strategies.

Liang Wenfeng, an AI aficionado, co-established High-Flyer in 2015. Reportedly, Wenfeng began experimenting with trading during his time as a student at Zhejiang University. He then launched High-Flyer Capital Management as a hedge fund in 2019, concentrating on the creation and implementation of AI algorithms.

In 2023, High-Flyer initiated DeepSeek as a specialized laboratory dedicated to the exploration of AI tools, functioning independently from its core financial operations. Subsequently, with High-Flyer as a key investor, the lab evolved into a separate entity, also named DeepSeek.

From its inception, DeepSeek prioritized constructing its own data center clusters for model training. However, similar to other AI firms in China, DeepSeek has experienced repercussions from U.S. export restrictions on hardware. To facilitate the training of one of its more recent models, the company was compelled to utilize Nvidia H800 chips, a less potent iteration of the H100 chip available to U.S. counterparts.

DeepSeek’s technical personnel are reportedly composed of predominantly young individuals. The company is known to actively recruit doctoral-level AI researchers from prominent Chinese universities. Additionally, according to reports in The New York Times, DeepSeek also employs individuals without formal computer science backgrounds to enhance its technology’s comprehension across diverse subject areas.

Robust AI Models from DeepSeek

DeepSeek introduced its initial suite of models—DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat—in November 2023. Nevertheless, it was not until the subsequent spring, with the unveiling of its next-generation DeepSeek-V2 family of models, that the broader AI industry began to recognize its significance.

DeepSeek-V2, a versatile system capable of text and image analysis, demonstrated strong performance across various AI benchmarks. Notably, it offered considerably more cost-effective operation compared to contemporary models at that time. This development spurred DeepSeek’s domestic competitors, including ByteDance and Alibaba, to decrease usage costs for some of their models and offer others entirely without charge.

The introduction of DeepSeek-V3 in December 2024 further solidified DeepSeek’s growing prominence.

DeepSeek’s internal benchmark assessments indicate that DeepSeek V3 surpasses both downloadable, openly available models like Meta’s Llama and “closed” models accessible exclusively through an API, such as OpenAI’s GPT-4o.

DeepSeek’s R1 “reasoning” model is considered equally noteworthy. Launched in January, DeepSeek asserts that R1 rivals OpenAI’s o1 model in performance on crucial benchmarks.

As a reasoning model, R1 incorporates self-verification capabilities, effectively mitigating typical model inaccuracies. While reasoning models may require slightly longer processing times—ranging from seconds to minutes—to generate solutions compared to standard non-reasoning models, they offer enhanced reliability in fields such as physics, science, and mathematics.

However, DeepSeek’s AI models, including R1 and DeepSeek V3, are subject to certain constraints. Being developed in China, they undergo evaluation by China’s internet regulatory bodies to ensure their outputs align with “core socialist values.” For instance, within DeepSeek’s chatbot application, R1 is programmed to avoid responding to inquiries concerning Tiananmen Square or Taiwan’s autonomy.

In March, DeepSeek’s website traffic exceeded 16.5 million visits. According to David Carr, editor at Similarweb, “[F]or March, DeepSeek ranks second, despite experiencing a 25% decrease in traffic from February, based on daily visits.” Despite this impressive figure, it remains considerably smaller compared to ChatGPT, which surpassed 500 million weekly active users in March.

A Disruptive Strategy

DeepSeek’s precise business model remains somewhat ambiguous. The company prices its offerings significantly below prevailing market rates and provides certain services without charge. Furthermore, despite substantial interest from venture capital, DeepSeek is not currently seeking external investment.

DeepSeek attributes its extreme cost competitiveness to advancements in efficiency. However, certain industry experts have questioned the veracity of the figures provided by the company.

Regardless, developers have shown strong adoption of DeepSeek’s models. While not strictly open source in the traditional sense, they are available under licenses that permit commercial application. Clem Delangue, CEO of Hugging Face, a prominent platform hosting DeepSeek’s models, reports that developers on Hugging Face have generated over 500 “derivative” models based on R1, collectively accumulating 2.5 million downloads.

DeepSeek’s achievements against larger, more established competitors have been characterized as both “upending AI” and “over-hyped.” The company’s progress has been partially credited with an 18% decline in Nvidia’s stock price in January and prompted a public statement from OpenAI CEO Sam Altman. In March, U.S. Commerce Department bureaus informed staff that DeepSeek would be prohibited on their government-issued devices, as reported by Reuters.

Microsoft has announced the availability of DeepSeek on its Azure AI Foundry service, a platform consolidating AI services for enterprises. During Meta’s first-quarter earnings call, CEO Mark Zuckerberg, when questioned about DeepSeek’s influence on Meta’s AI expenditure, affirmed that investment in AI infrastructure would remain a “strategic advantage” for Meta. In March, OpenAI labeled DeepSeek as “state-subsidized” and “state-controlled,” recommending that the U.S. government consider imposing restrictions on DeepSeek’s models.

Jensen Huang, CEO of Nvidia, acknowledged DeepSeek’s “excellent innovation” during Nvidia’s fourth-quarter earnings call, noting that DeepSeek and other reasoning models are beneficial for Nvidia due to their substantial computational requirements.

Conversely, some entities, including entire nations and governments such as South Korea and New York state, are implementing bans on the use of DeepSeek on government devices.

The future trajectory of DeepSeek remains uncertain. Further advancements in its models are anticipated. However, the U.S. government is seemingly becoming increasingly cautious regarding perceived detrimental foreign influence. In March, The Wall Street Journal reported on the likely U.S. government ban of DeepSeek on government devices.