In Spain and Latin America, when we talk about technology, we tend to focus on the United States. However, in recent years, countries such as China and South Korea have cornered a large part of the global market for applications, services, and electronic devices. The same is true in the field of artificial intelligence. We are familiar with OpenAI, Meta, Google, Microsoft and Apple, all US companies. But we know little or nothing about what is happening in other countries. Without going any further, the announcement of DeepSeek, a Chinese artificial intelligence model, has taken many by surprise. A model that rivals OpenAI and Meta.
Recently, the President of the United States and several companies announced a multimillion-pound project focused on boosting the country’s artificial intelligence industry. Named the Stargate Project, one of the reasons for launching this project is China’s progress in AI projects. Without going any further, one of them has received widespread media attention. Its name is DeepSeek and, after a year of work, it has achieved an AI model that is on par with well-known models such as GPT (from OpenAI), Llama (from Meta) and Claude (from Anthropic). And this is only the beginning.
What is DeepSeek?
Let’s start at the beginning. DeepSeek is an artificial intelligence laboratory based in Hangzhou, China. Interestingly, this booming city is also home to major technology companies such as Alibaba, NetEase, Geely and HikVision. The laboratory was founded in May 2023 by Liang Wenfeng. It is part of a larger company, High Flyer, which was also created by Wenfeng himself. Back in 2016, High Flyer was created as a hedge fund, i.e. an investment fund, which would eventually be worth more than $15 billion. One of its most ambitious projects has been Fire Flyer, a subsidiary focused on research into deep learning, a branch of artificial intelligence.
So, through Fire Flyer, Liang Wenfeng created supercomputers that were initially intended to process financial data. But in May 2023, he decided to change his strategy and create DeepSeek, a project to design his own artificial intelligence model. He had the money and the necessary infrastructure. So, in November of that same year, DeepSeek Coder, his first AI model, was already available. He made it available to everyone, free of charge and open source under an MIT Licence.
That same month, it released its second model, DeepSeek LLM, with its own chatbot or conversational bot. Its own ChatGPT, in other words. And with a processing capacity of 67 billion parameters. That is, 67B. To give us an idea, OpenAI’s GPT-1 (2018) was capable of processing 117 million parameters. And GPT-3 (2020), 175 billion parameters.
Why everyone’s talking about it?
In May last year, this unique AI laboratory announced DeepSeek V2. Capable of rivalling the competition – and, what’s more, it stood out for being cheaper and more efficient. In December of the same year, DeepSeek V3 was announced. With 671 billion parameters. And training in just 55 days at a much lower cost than the competition. In tests, its figures exceeded those of Llama 3.1 (from Meta), GPT-4o (from OpenAI) and Claude 3.5 Sonnet (from Anthropic). With fewer resources, at lower cost and in less time, DeepSeek had achieved an artificial intelligence model capable of rivalling the giants of Silicon Valley.
At the same time, it is also developing a model called DeepSeek R1. Trained for logic, mathematical reasoning and real-time problem solving. Its processing capacity is 671 billion parameters. It also has a reduced version, R1 Zero, with 37 billion parameters. And, according to tests carried out, it is on a par with the AI models from OpenAI and Meta, among others.
Another peculiarity of DeepSeek is that it does not depend financially on Chinese companies such as Baidu, Alibaba or ByteDance. On the contrary. It is a worthy rival. And its AI model is also serious competition for the models of these Chinese internet giants. So sooner rather than later, DeepSeek was bound to go global, as it has recently done.
And, not long ago, it became known in the United States. A market dominated by its own AI assistants. Thus, this Chinese proposal has decided to release its own conversational bot, its own ChatGPT that can be installed for free on iPhone and Android. At the time of writing, the DeepSeek iPhone app is the most downloaded. ChatGPT ranks third. And on Google Play, it is the 20th most downloaded app. And climbing the ranks.
The keys to DeepSeek’s success
DeepSeek has all the ingredients to become one of the leading contenders in global artificial intelligence industry. For starters, it is not seeking financial gain. It is offered free of charge, with open source code for developers, and costs less than rival AI models from OpenAI, Google and Microsoft, which are forced to offer paid options to recoup their investment, cover the infrastructure costs involved in using artificial intelligence, and ultimately turn a profit.
According to research by Epoch AI, DeepSeek V3 requires ten times less computing power than Meta’s Llama 3.1. One reason is that, as Wired explains, ‘it has made significant progress in Multi-head Latent Attention (MLA) and Mixture-of-Experts, two technical designs that make DeepSeek models more cost-effective by requiring fewer computing resources to train’. The research explains at length what both technical approaches consist of. For those wanting to delve deeper.
How to start using DeepSeek AI
DeepSeek, which gives its name to the company that develops this AI model and to the AI itself, is available in several versions. To get started, you can use the web version, as with ChatGPT or Gemini. Simply register as a user and start interacting with DeepSeek from your web browser. And as we saw earlier, DeepSeek has recently become available as a virtual assistant for iPhone and Android. It can be downloaded on iOS and Android.
For those who want to work with this artificial intelligence and develop their own apps and services, a paid API and development platform are available. It also has its own space on GitHub, where it explains how to test DeepSeek locally or from external services such as Hugging Face.
