Chinese artificial intelligence start-up DeepSeek said a significant upgrade to its R1 reasoning model has given it performance comparable with the latest Western models from the likes of OpenAI and Google.

It said the R1-0528 release, the first major update to the R1 “reasoning” model released in January, beats the performance of Alibaba’s Qwen3, which was released a month earlier, in late April.

The start-up said the R1 update improves its reasoning and creative writing abilities and coding output.

Reliability

It also achieved a 50 percent reduction in “hallucinations”, or the generation of incorrect or misleading information.

The improved capabilities were largely the result of adding more computing resources at the post-training stage, the company said.

Benchmark tests including maths, coding and general logic showed that R1-0528 surpassed domestic Chinese models and attained similar levels to global models including OpenAI’s o3 and Google’s Gemini2.5-Pro, DeepSeek said.

The start-up’s models have attracted attention in part due to the high performance they have achieved, while requiring far less computing power than other major models to build and operate.

DeepSeek’s models are also open source, which contrasts with the closed-source method employed by OpenAI and some others.

Tencent, Baidu and ByteDance all said they would integrate R1-0528 into their cloud platforms for corporate clients.

DeepSeek also said it had distilled knowledge from the new model to create a smaller one that matches the performance of Alibaba’s Qwen3-235B while being almost 30 times smaller in parameter size.

Open-source approach

The company said its experiment could be of interest for academic research into reasoning models or the commercial development of lighter-weight models.

DeepSeek recently released a series of open-source projects on GitHub revealing details about how it trained its R1 and V3 models.

The eight open-source projects were the first time it had disclosed details of the techniques it used to gain optimal performance from compute, communications and storage, three key aspects of model training.

The company’s developers, who are mostly young university graduates, said they were disclosing the company’s “battle-tested building blocks” to share “our small-but-sincere progress with full transparency”.