The Chinese company, DeepSeek, today offers a learning solution for AI that is much less expensive than its American competitor OpenAI.
L’artificial intelligence is a vast subject, covering different themes, ranging from practical applications to theoretical aspects, including the learning of data from these AIs through what we call LLM (language models). Today, the companies behind many AI systems are attracting concerns as to the economic and environmental costs they generate. DeepSeek offers a solution to this problem, at least from an economic point of view.
Faster and Cheaper
By progressing and becoming more and more efficient, artificial intelligence increases the parameters of its language models to stay relevant and understand increasingly complex queries. An improvement that comes at a cost. As an example, version 3 of ChatGPTwhich would count 175 billion parameters according to The Vergecost several hundred million dollars to develop. OpenAI, its parent company, is also regularly singled out for the astronomical sums that allow it to stay afloat.
According to Jim Fan, research scientist at NvidiaDeepSeek, based in Hangzhou, China, is considered “the biggest outsider” of 2025 in the field of large open source LLMs, reports the Chinese media South China Morning Post.
Compared to ChatGPT, DeepSeek V3which has 671 billion parameters, was trained in just two months at a cost of 5,58 millions of American dollars.
An Open Source model
For its part, in addition to being economical, DeepSeek positions itself as a model open source which seems to prove itself among professionals. “DeepSeek V3 appears to be a more powerful model, with only 2.8 million GPU hours”said Andrej Karpathy, former head of AI at Tesla and a founding member of the OpenAI team, in a post on X (Twitter). For comparison, Llama 3.1the Meta group’s AI, required 30.8 million GPU hours of training, according to South China Morning Post.
“I’ve been following DeepSeek for a long time. Last year they had one of the best open coding models »says Fan. He goes on to say that these open source models will put enormous pressure on the sector which will lead to rapid evolution.
For the most curious, the DeepSeek model is accessible on Hugging Face and it is accompanied by its documentation on GitHub.
Did you know? Google News lets you choose your media. Don’t miss Frandroid and Numerama.
