Deepseek went viral.
Chinese Ai Lab Deepseek broke the basic awareness this week after Its Chatbot app rose to the top of Apple App Store charts (And Google Play, as well). AI's AI models of Deptseek, trained using compute-effective techniques, has led to Wall Street analysts – and technologists – To ask if the US can maintain its lead in the AI breed and if the demand for AI chips will maintain.
But where did the Deepseek come from, and how did it rise quickly to international popularity?
Deepseek's Businessman's Source
Deepseek is supported by high-flyer capital management, a Chinese hedge fund that uses AI to inform its trading decisions.
Ai see Liang wenfeng The co-established high-flyer in 2015. Wenfeng, who was reported to have begun to dabbling trading while a student at Zhejiang University, launched high-flyer capital management as a fence fund in 2019 dedicated to developing and eliminating AI algorithms.
In 2023, High-Flyer started Deepseek as a lab dedicated to researching AI tools separate from the financial business. In high-flyer as one of its investors, the lab has infiltrated its own company, also called Deepseek.
From one day, Deepseek built its own clusters of data center for model training. But like other AI companies in China, DeepSeek has been affected by US exporting restrictions on hardware. 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.
The Technical Team Team Team is said to be a child -KWW. The company reported aggressive recruitment Doctorate AI researchers from the leading Chinese universities. Deepseek also hired people without any background in computer science To help tech better understand a wide range of topics, every New York Times.
The strong depsek models
Deepseek announced the first set of models-depseek coder, Deptseek LLM, and Deepseek Chat-in November 2023. But not until the last spring, when the startup released the next Gen DeviceEk-V2 family of models, the AI industry began to notice.
Deepseek-V2, a general-purpose text- and image analysis system, has performed well on various Ai-and-cheaper benchmarks to run than comparable models at the time. It forced Deepseek's domestic competition, including bytedance and alibaba, to cut use prices for some of their models, and make others completely free.
DeepSeek-V3Launched in December 2024, added only to Deepseek's being known.
According to DeepSeek's internal testing benchmark, DeepSeek V3 Outperforms are both downloadable, openly available models like Meta's Llama and “closed” models that can only be accessible by an api, such as Openai's GPT-4O.
Equal wonder is the “Reason” model of the Deptseek. Released in January, the claims to Deepseek R1 performs as well as Openai model of O1 on major benchmarks.
As a model of reasoning, the R1 is an effective fact-check itself, helping to avoid some of the pitfalls that usually travel the models. Reasoning models take a little-standard seconds to minutes longer-to get to solutions compared to a typical non-compelling model. The upside is that they tend to be more reliable in domains such as physics, science, and mathematics.
There is a downside to R1, Deepseek V3, and other depsek models, however. As a developed AI developed by Chinese, they are subject to Benchmarking Through China's Internet regulator to ensure its responses are “embody core socialist value.” In DeepSeek's Chatbot app, for example, R1 will not answer questions about Tiananmen Square or Taiwanese autonomy.
A disturbing approach
If Deepseek has a business model, it's unclear what that model is, exactly. The price of its company is well with the products and services below the market value – and gives others free. It also doesn't take the investor's moneyDespite a ton of interest of VC.
The way it tells Deepseek, excellence breakthroughs enables to maintain extreme cost competition. Some experts dispute The figures the company provides, however.
Regardless of the case, the developers are taken to Deepseek models, which are not open resources because the phrase is usually understood but available under the allowable licenses that provide commercial use. According to Clem Delangue, the CEO of the Hugging Face, one of the platforms that shown deptseek models, Those who develop in the hugging face created more than 500 “derivative” R1 models That has raised up 2.5 million downloads combined.
Deepseek's success against the larger and more established rivals has been Described as “Upending AI” and “Over-hyped.” Company success is at least part responsible for Causes Nvidia's stock price collapse of 18% in January, and for The choice of a public response From Openai CEO Sam Altman. In March, the US Commerce Department Bureaus told staff that Deepseek is prohibited on their government devicesAccording to Reuters.
Microsoft announced that DeepSeek is available to its Azure AI Foundry serviceMicrosoft platform combining -Includes AI services for businesses under a single banner. When asked about Deepseek's impact on Meta's AI expenditure during the first quarter revenue call, CEO Mark Zuckerberg said Spending on AI infrastructure will continue to be a “strategic advantage” For the meta. In March, Openai called the deptseek “state-subsidized” and “state controlled,” And it is recommended that the US government consider the ban on the deptseek models.
During Nvidia's fourth quarter call, CEO Jensen Huang emphasized “great change,” Saying this and other “reasoning” models are great for Nvidia because they need more compute.
At the same time, Some companies prohibit depseekand so is the whole countries and Governments, including South Korea. New York State also Deepseek banned from using government devices.
When it comes to what the Deptseek can handle, it's not clear. Enhanced models are provided. But the US government will appear to be growing carefully with what it sees as a harmful influence on the foreigner. In March, The Wall Street Journal reported that The US is likely to prohibit depseek with government devices.
This story was originally published on January 28, 2025, and was regularly updated.
