China's DeepSeek releases reasoning model R1, free for anyone to download

HANGZHOU, JANUARY 2025 - The Chinese AI company DeepSeek released the model R1 on 20 January. Like OpenAI's model o1, R1 thinks first and only then gives its answer. According to DeepSeek, it performs about as well as o1 on difficult maths and programming questions. What makes it special is that anyone may download, use and adapt the model, including for commercial purposes. DeepSeek also offers a free app for iOS and Android.

What a reasoning model is

An ordinary language model such as ChatGPT writes its answer straight away. A reasoning model first writes out a long line of thought, step by step, and only then gives the answer. OpenAI showed this in September with o1, but that model can only be used through OpenAI itself, and its line of thought stays hidden. With R1, anyone can run the model on their own computers.

How R1 was trained

That a model learns to reason is less obvious than it seems. You can give it examples of people reasoning aloud, but these are rare and expensive to make. DeepSeek chose a different route. The researchers let the model solve problems whose answers can be checked automatically, such as sums and programming tasks. If the result was correct, the model got a reward, and if not, it did not. This is called reinforcement learning. According to the researchers, the model in this way learned reasoning behaviour by itself, such as checking its own work and trying a different approach. The first trial version, R1-Zero, received no human examples at all. Both models are built on the same DeepSeek base model, V3-Base.

What does "open model" mean?

An AI model is an enormous file of numbers, called the weights, which hold everything it has learned. DeepSeek publishes these weights under the MIT licence, one of the most permissive licences there is. Anyone can download them. The data the model was trained on is not shared.

R1 is large: 671 billion parameters, of which the model uses only 37 billion for each word. This is possible because it consists of many specialised parts, of which only a few are at work at any time. That makes it faster and more economical than a model of the same size that always uses everything. For people with less computing power, DeepSeek also published six smaller versions, from 1.5 to 70 billion parameters. These were trained on texts that R1 itself wrote. According to the company, the version with 32 billion parameters scores better than OpenAI's small model o1-mini.

What the figures say

According to DeepSeek, R1 scores 79.8 per cent on the maths test AIME, against 79.2 per cent for o1. On the maths test MATH-500 it is 97.3 against 96.4 per cent. In programming contests on Codeforces, o1 is slightly ahead, with a rating of 2,061 against 2,029 points. These figures come from the company itself, and independent tests still have to confirm them.

Looking back from 2026

On 27 January 2025, DeepSeek's app was the most downloaded free app in the American App Store. According to Wikipedia, Nvidia's share price fell by about 18 per cent that day, because investors wondered whether strong AI really needs so many expensive chips. DeepSeek published the method in a scientific paper, which later appeared in Nature.

R1 showed that the approach of o1 did not belong to one company alone. An open model means that others can use it, adapt it and make it smaller. DeepSeek keeps sharing its research: at the end of October 2026, the company released the code of an AI assistant that takes on programming work independently, still as a trial version.