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

DeepSeek has gone viral.

Chinese AI laboratory DeepSeek broke into the mainstream consciousness this week after its chatbot app roseate to the apical of the Apple App Store charts (and Google Play, arsenic well). DeepSeek’s AI models, which were trained utilizing compute-efficient techniques, have led Wall Street analysts — and technologists — to question whether the U.S. tin support its pb successful the AI contention and whether the request for AI chips volition sustain.

But wherever did DeepSeek travel from, and however did it emergence to planetary fame truthful quickly?

DeepSeek’s trader origins

DeepSeek is backed by High-Flyer Capital Management, a Chinese quantitative hedge money that uses AI to pass its trading decisions.

AI enthusiast Liang Wenfeng co-founded High-Flyer successful 2015. Wenfeng, who reportedly began dabbling successful trading portion a pupil astatine Zhejiang University, launched High-Flyer Capital Management arsenic a hedge money successful 2019 focused connected processing and deploying AI algorithms.

In 2023, High-Flyer started DeepSeek arsenic a laboratory dedicated to researching AI tools abstracted from its fiscal business. With High-Flyer arsenic 1 of its investors, the laboratory spun disconnected into its ain company, besides called DeepSeek.

From time one, DeepSeek built its ain information halfway clusters for exemplary training. But similar different AI companies successful China, DeepSeek has been affected by U.S. export bans connected hardware. To bid 1 of its much caller models, the institution was forced to usage Nvidia H800 chips, a less-powerful mentation of a chip, the H100, disposable to U.S. companies.

DeepSeek’s method squad is said to skew young. The institution reportedly aggressively recruits doctorate AI researchers from apical Chinese universities. DeepSeek besides hires radical without immoderate machine subject background to assistance its tech amended recognize a wide scope of subjects, per The New York Times.

DeepSeek’s beardown models

DeepSeek unveiled its archetypal acceptable of models — DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat — successful November 2023. But it wasn’t until past spring, erstwhile the startup released its next-gen DeepSeek-V2 household of models, that the AI manufacture started to instrumentality notice.

DeepSeek-V2, a general-purpose text- and image-analyzing system, performed good successful assorted AI benchmarks — and was acold cheaper to tally than comparable models astatine the time. It forced DeepSeek’s home competition, including ByteDance and Alibaba, to chopped the usage prices for immoderate of their models, and marque others wholly free.

DeepSeek-V3, launched successful December 2024, lone added to DeepSeek’s notoriety.

According to DeepSeek’s interior benchmark testing, DeepSeek V3 outperforms some downloadable, openly disposable models similar Meta’s Llama and “closed” models that tin lone beryllium accessed done an API, similar OpenAI’s GPT-4o.

Equally awesome is DeepSeek’s R1 “reasoning” model. Released successful January, DeepSeek claims R1 performs arsenic good arsenic OpenAI’s o1 model connected cardinal benchmarks.

Being a reasoning model, R1 efficaciously fact-checks itself, which helps it to debar immoderate of the pitfalls that usually travel up models. Reasoning models instrumentality a small longer — usually seconds to minutes longer — to get astatine solutions compared to a emblematic non-reasoning model. The upside is that they thin to beryllium much reliable successful domains specified arsenic physics, science, and math.

There is simply a downside to R1, DeepSeek V3, and DeepSeek’s different models, however. Being Chinese-developed AI, they’re taxable to benchmarking by China’s net regulator to guarantee that its responses “embody halfway socialist values.” In DeepSeek’s chatbot app, for example, R1 won’t reply questions astir Tiananmen Square oregon Taiwan’s autonomy.

In March, DeepSeek surpassed 16.5 cardinal visits. “[F]or March, DeepSeek is successful 2nd place, contempt seeing postulation driblet 25% from wherever it was successful February, based connected regular visits,” David Carr, exertion astatine Similarweb, told TechCrunch. It inactive pales successful examination to ChatGPT, which surged past 500 cardinal play progressive users successful March.

A disruptive approach

If DeepSeek has a concern model, it’s not wide what that exemplary is, exactly. The institution prices its products and services good beneath marketplace worth — and gives others distant for free. It’s besides not taking capitalist money, contempt a ton of VC interest.

The mode DeepSeek tells it, ratio breakthroughs person enabled it to support utmost outgo competitiveness. Some experts dispute the figures the institution has supplied, however.

Whatever the lawsuit whitethorn be, developers person taken to DeepSeek’s models, which aren’t unfastened root arsenic the operation is commonly understood but are disposable nether permissive licenses that let for commercialized use. According to Clem Delangue, the CEO of Hugging Face, 1 of the platforms hosting DeepSeek’s models, developers connected Hugging Face person created implicit 500 “derivative” models of R1 that person racked up 2.5 cardinal downloads combined.

DeepSeek’s occurrence against larger and much established rivals has been described arsenic “upending AI” and “over-hyped.” The company’s occurrence was astatine slightest successful portion liable for causing Nvidia’s banal terms to driblet by 18% successful January, and for eliciting a nationalist response from OpenAI CEO Sam Altman. In March, U.S. Commerce section bureaus told staffers that DeepSeek volition beryllium banned connected their authorities devices, according to Reuters.

Microsoft announced that DeepSeek is disposable connected its Azure AI Foundry service, Microsoft’s level that brings unneurotic AI services for enterprises nether a azygous banner. When asked astir DeepSeek’s interaction connected Meta’s AI spending during its first-quarter net call, CEO Mark Zuckerberg said spending connected AI infrastructure volition proceed to beryllium a “strategic advantage” for Meta. In March, OpenAI called DeepSeek “state-subsidized” and “state-controlled,” and recommends that the U.S. authorities see banning models from DeepSeek.

During Nvidia’s fourth-quarter net call, CEO Jensen Huang emphasized DeepSeek’s “excellent innovation,” saying that it and different “reasoning” models are large for Nvidia due to the fact that they request truthful overmuch much compute.

At the aforesaid time, some companies are banning DeepSeek, and truthful are full countries and governments, including South Korea. New York authorities besides banned DeepSeek from being utilized connected authorities devices.

In May, Microsoft Vice Chairman and President Brad Smith said successful a Senate proceeding that Microsoft employees aren’t allowed to usage DeepSeek owed to information information and propaganda concerns.

As for what DeepSeek’s aboriginal mightiness hold, it’s not clear. Improved models are a given. But the U.S. authorities appears to beryllium growing wary of what it perceives arsenic harmful overseas influence. In March, The Wall Street Journal reported that the U.S. volition apt prohibition DeepSeek connected authorities devices.

This communicative was primitively published January 28, 2025, and volition beryllium updated regularly.