Chinese artificial intelligence (AI) start-up DeepSeek has released an update to its R1 logical reasoning model. This increases competitive pressure on US companies such as OpenAI, Reuters reported today.
The Chinese model DeepSeek R1 has quietly surprised everyone
DeepSeek launched its new model on the Hugging Face developer platform. However, it has not yet made a public announcement about the new model, nor has it released a description or comparison.
In the LiveCodeBench ranking, which compares the performance of AI models in programming and was developed by researchers from the University of California, Berkeley, Massachusetts Institute of Technology (MIT) and Cornell University, the updated R1 reasoning model ranked only slightly behind OpenAI’s o4 mini and o3 models in code generation, ahead of xAI’s Grok 3 mini and Alibaba’s Qwen 3 models. This is a significant achievement for the Chinese AI sector, as it shows that Chinese models are on par with the world’s best in the demanding discipline of program code generation.
DeepSeek is changing the rules of the game
DeepSeek refuted the belief that US export controls were hampering Chinese progress in AI earlier this year when it released models that were the same or better than the US’s best at a fraction of the price. The launch of the R1 model caused a sharp decline in technology stocks outside China in January and challenged the view that the expansion of AI requires enormous computing power and investment. DeepSeek is still widely expected to release R2, the successor to R1. Reuters reported in March, citing sources, that the release of R2 was originally planned for May. The company also released an update to its large language model V3 in March.
The primary purpose of language models is to process and generate text. Their strengths include fluent and natural language, and their capabilities include conversation, translation, summarization, and style. Logical reasoning models focus primarily on problem solving and precise logic. Their strengths include correct and systematic reasoning, with less emphasis on language. It is not just a matter of the answer sounding natural, but of it being logically correct and well thought out, and of being able to solve a mathematical problem or program functional code, for example.
Source: Reuters
