How a small Chinese AI company is shaking up US tech heavyweights

Chinese artificial intelligence (AI) company DeepSeek has sent shockwaves through the tech community^[1], with the release of extremely efficient AI models that can compete with cutting-edge products from US companies such as OpenAI and Anthropic.

Founded in 2023, DeepSeek has achieved its results^[2] with a fraction of the cash and computing power of its competitors.

DeepSeek’s “reasoning” R1 model, released last week, provoked excitement among researchers, shock among investors, and responses from AI heavyweights. The company followed up on January 28 with a model^[3] that can work with images as well as text.

So what has DeepSeek done, and how did it do it?

What DeepSeek did

In December, DeepSeek released its V3 model^[4]. This is a very powerful “standard” large language model that performs at a similar level to OpenAI’s GPT-4o and Anthropic’s Claude 3.5.

While these models are prone to errors and sometimes make up their own facts^[5], they can carry out tasks such as answering questions, writing essays and generating computer code. On some tests^[6] of problem-solving and mathematical reasoning, they score better than the average human.

V3 was trained at a reported cost^[7] of about US$5.58 million. This is dramatically cheaper than GPT-4, for example, which cost more than US$100 million^[8] to develop.

DeepSeek also claims to have trained V3 using around 2,000 specialised computer chips, specifically H800 GPUs made by NVIDIA^[9]. This is again much fewer than other companies, which may have used up to 16,000^[10] of the more powerful H100 chips.

On January 20, DeepSeek released another model, called R1^[11]. This is a so-called “reasoning” model, which tries to work through complex problems step by step. These models seem to be better at many tasks that require context and have multiple interrelated parts, such as reading comprehension and strategic planning.

The R1 model is a tweaked version of V3, modified with a technique called reinforcement learning. R1 appears to work at a similar level to OpenAI’s o1^[12], released last year.

DeepSeek also used the same technique to make “reasoning” versions of small open-source models that can run on home computers.

This release has sparked a huge surge of interest in DeepSeek, driving up the popularity of its V3-powered chatbot app^[13] and triggering a massive price crash^[14] in tech stocks as investors re-evaluate the AI industry. At the time of writing, chipmaker NVIDIA has lost around US$600 billion^[15] in value.

How DeepSeek did it

DeepSeek’s breakthroughs have been in achieving greater efficiency: getting good results with fewer resources. In particular, DeepSeek’s developers have pioneered two techniques that may be adopted by AI researchers more broadly.

The first has to do with a mathematical idea called “sparsity”. AI models have a lot of parameters that determine their responses to inputs (V3 has around 671 billion), but only a small fraction of these parameters is used for any given input.

However, predicting which parameters will be needed isn’t easy. DeepSeek used a new technique to do this, and then trained only those parameters. As a result, its models needed far less training than a conventional approach.

The other trick has to do with how V3 stores information in computer memory. DeepSeek has found a clever way to compress the relevant data, so it is easier to store and access quickly.