DeepSeek’s R1-0528 now ranks right behind OpenAI's o4-mini

DeepSeek has rolled out R1-0528, a major upgrade to the Chinese start-up’s R1 reasoning model, which was released in January. The upgrade arrived just one month after Alibaba Group Holding’s Qwen3 beat the original DeepSeek R1 in LiveBench, an online benchmark for open-source artificial-intelligence models.

DeepSeek’s upgraded R1-0528 model now stands alongside leading AI models from OpenAI and Google in performance. The comeback shows how quickly China’s big technology firms and newer tech firms are pushing to improve their AI tools.

In its statement, DeepSeek said R1-0528 shows better reasoning and creative writing skills. The update also brings stronger coding ability. Most importantly, the company claims the model now produces 50% fewer “hallucinations.”

DeepSeek explained that the upgrades came from extra computing power invested during the post-training phase, when engineers fine-tune a model after the main training process. During the post-training phase, engineers aim to increase the model’s efficiency and enhance its accuracy and safety.

R1-0528 now ranks right behind OpenAI’s o3 and o4-mini

On LiveCodeBench, which measures AI model performance, R1-0528 now ranks just behind OpenAI’s o4-mini and o3 models.

“DeepSeek’s latest upgrade is sharper on reasoning, stronger on math and code, and closing in on top-tier models like Gemini and O3,” said Adina Yakefu, an AI researcher at Hugging Face.

She added that the new version shows “major improvements in inference and hallucination reduction” and proves the start-up is not merely catching up but actively competing.

The rapid progress came after Washington had restricted advanced chips and other technology exports to China. Yet Chinese firms continue to refine their systems. Earlier this month, Baidu and Tencent described ways they are making their models run more efficiently despite limited access to cutting-edge semiconductors.

Nvidia chief executive Jensen Huang criticized export controls on Wednesday. “The U.S. has based its policy on the assumption that China cannot make AI chips,” he said. “That assumption was always questionable, and now it’s clearly wrong. The question is not whether China will have AI. It already does.”

DeepSeek raised the performance of Alibaba’s Qwen3 8B model by 10%

DeepSeek also said it distilled the reasoning steps used in R1-0528 into Alibaba’s Qwen3 8B Base model. That process created a new, smaller model that surpassed Qwen3’s performance by more than 10%, according to the company. At the same time, the model was 30 times smaller.

“We believe the chain-of-thought from DeepSeek-R1-0528 will hold significant importance for academic research on reasoning models and industrial work on small models,” the firm stated.

According to Reuters, a DeepSeek representative told a WeChat group that the change was a “minor trial upgrade” that was already open for public testing. In response to fiercer competition, Google has discounted some Gemini access tiers, while OpenAI introduced the lower-cost o3 Mini model.

Cryptopolitan Academy: Coming Soon - A New Way to Earn Passive Income with DeFi in 2025. Learn More