Deepseek has gone viral.
The Chinese AI Deepseek workshop broke the usual consciousness this week after the Chatbot application at the top of the Apple App Store charts (and Google Play, too). Deepseek AI models trained using efficient calculation techniques, led Wall Street analysts-and technologists-to dispute whether the US can maintain its lead in the AI race and if the demand for AI chip will be maintained.
But where did Deepseek come from and how did it get to international reputation so fast?
DEEPSEEK TRADER ORIGINS
Deepseek is supported by high quality capital management, a Chinese quantitative hedge fund that uses AI to inform its commercial decisions.
The enthusiastic Liang Wenfeng founded High-Flyer in 2015. Wenfeng, who reportedly started trading, while a student at Zhejiang University started High-Flyer Capital Management as a 2019 Risk Functional Chapter.
In 2023, High-Flyer started Deepseek as a laboratory dedicated to AI tool research separately from his financial activities. With High-Flyer as one of its investors, the lab began in his company, also called Deepseek.
From the first day, Deepseek has created its own modeling data centers. But like other AI companies in China, Deepseek has been influenced by US export bans on material. To train one of its latest models, the company was forced to use the Nvidia H800 chips, a less powerful version of a chip, the H100, available to US companies.
Deepseek’s technical team is said to violate young people. The company Reportedly he hires aggressively AI doctoral researchers from top Chinese universities. Deepseek also hires people with no computer background To help her technology better understand a wide range of issues, according to the New York Times.
Deepseek’s powerful models
Deepseek revealed its first set of models-deeseseek coder, Deepseek LLM and Deepseek Chat-in November 2023, but was not until last spring, when the launch was released by the DEEPSEEK-V2 family, which began to know the AI industry.
Deepseek-V2, a general-purpose text analysis system, is performed well at various AI reference points-and it was much cheaper to run than comparable models at that time. Deepseek’s domestic competition was forced, including byTedance and Alibaba, to reduce usage prices for some of their models and make others completely free.
Deepseek-V3, which began in December 2024, was only added to the reputation of Deepseek.
According to Deepseek’s internal reference tests, the Deepseek V3 exceeds the two models that can download, openly available, such as Meta’s Llama and “closed” models that can only be approached through an API, such as Openai’s GPT-4O.
Equally impressive is the Deepseek R1 “Reasoning” model. Released in January, Deepseek claims that the R1 runs as well as the O1 OPENAI model at key benchmarks.
Being a model of reasoning, the R1 effectively controls the same, which helps it avoid some of the traps that normally travel to models. Logic models last a little longer-usually seconds to minutes-to reach solutions compared to a typical non-hedgean model. The rise is that they tend to be more reliable in areas such as physics, science and mathematics.
But there is a downside to R1, Deepseek V3 and other Deepseek models. Being developed by Chinese AI, subject to performance rating from the Chinese regulatory authority to ensure that her answers “incorporate the core socialist values”. In Deepseek’s Chatbot application, for example, R1 will not answer questions about Tiananmen Square or Taiwan’s autonomy.
An annoying approach
If Deepseek has a business model, it’s not clear what this model is, exactly. The company invoice its products and services well below market value – and gives others away for free. It also does not receive investor money, despite the tone of interest in VC.
The way Deepseek says it, its performance discoveries allowed it to maintain the competitiveness of extreme costs. Some experts conflict However, the company’s details were delivered.
Whatever the case, developers have received Deepseek models, which are not open, as the phrase is usually understood but available with permissible licenses that allow commercial use. According to Clem Delangue, Hugging Face’s chief executive, one of the platforms that host Deepseek models, Developers in Hugging Face have created over 500 “derivatives” R1 models which have raised 2.5 million shots in combination.
Deepseek’s success against the biggest and most established opponents were described as “UPENDING AI” and “excessively.” The company’s success was at least partly responsible for the fall of NVIDIA’s stock price by 18% in January and for causes a public response by Openai Sam Altman’s chief executive. In March, the offices of the US Department of Commerce reported to personal Deepseek will be banned in their government devicesAccording to Reuters.
Microsoft has announced that Deepseek is available at the Azure Ai Foundry service, the Microsoft platform that collects AI services for only one banner. When asked about Deepseek’s impact on META’s AI expenses during the first quarter profits, Managing Director Mark Zuckerberg said that AI infrastructure costs would continue to be a “strategic advantage” for META. In March, Openai called on Deepseek “state subsidy” and “state -controlled” and recommends that the US government is considering banning Deepseek models.
During the profits of Nvidia’s fourth quarter, Managing Director Jensen Huang underlined Deepseek’s “excellent innovation”, saying that other “logic” models are great for Nvidia because they need much more calculated.
At the same time, some companies prohibit Deepseek, and so they are whole countries and governments, including South Korea. New York State as well banned Deepseek to be used on government devices.
As for what the future of Deepseek can keep, it is not clear. Improved models are given. But the US government seems to be Greater cautious for what he perceives as harmful foreign influence. In March, Wall Street Journal reported that US will most likely ban Deepseek in government devices.
This story was first published on January 28, 2025 and will be updated regularly.
