China’s DeepSeek quietly updates its R1 reasoning model, narrowing the AI race gap with U.S. giants like OpenAI and outperforming rivals like Grok 3 and Qwen 3.
 |
| DeepSeek’s latest upgrade to its R1 reasoning model boosts performance in coding benchmarks, signaling China’s rapid advance in the global AI arms race. Image: CH |
Beijing, China – May 30, 2025:
Chinese artificial intelligence startup DeepSeek has quietly released an upgraded version of its flagship R1 reasoning model, underscoring the country's accelerating push in the global AI race. The update, known as R1-0528, was published early Thursday on the open developer platform Hugging Face without fanfare, lacking a formal company announcement or detailed release notes.
Despite the subdued launch, the implications are significant. According to the LiveCodeBench leaderboard—an AI benchmark maintained by researchers from UC Berkeley, MIT, and Cornell—the upgraded DeepSeek model now ranks just behind OpenAI’s o4 mini and o3 reasoning models for code generation. Notably, it surpasses competitors such as Elon Musk’s xAI Grok 3 mini and Alibaba’s Qwen 3, reinforcing DeepSeek’s standing as one of China’s most formidable AI developers.
The news was first reported by Bloomberg on Wednesday, which cited a DeepSeek representative informing a WeChat user group that the upgrade was a "minor trial" and that testing had already begun. While the company has not yet provided technical documentation or performance comparisons, third-party benchmarks suggest a tangible leap in capability for a model that had already disrupted expectations earlier this year.
The original R1 model, launched in January, sparked global attention by delivering performance comparable to or better than some leading U.S. models—despite operating with significantly lower compute resources. The release rattled markets and helped challenge the prevailing assumption that China's AI sector was being stifled by U.S. export controls on advanced chips.
DeepSeek’s momentum has pressured global competitors. U.S. companies like Google and OpenAI have since rolled out more affordable AI tiers, including OpenAI’s o3 Mini, designed to deliver competitive reasoning abilities with less computational overhead. Meanwhile, Chinese tech giants including Alibaba and Tencent have introduced new models claiming superiority over DeepSeek’s offerings, though benchmark results remain mixed.
The release of R1-0528 also comes amid mounting anticipation for R2, a full successor to the R1 line. According to a Reuters report in March, the R2 model was initially scheduled for release in May, although no further updates have been provided. In the meantime, DeepSeek has been active, also upgrading its V3 large language model in March to keep pace with evolving standards in natural language processing and reasoning.
The latest development cements DeepSeek’s role as a major player in the AI landscape and highlights how Chinese firms are leveraging efficient architectures and strategic releases to stay in step with, or even challenge, their Western counterparts—despite restricted access to high-end GPUs and cloud infrastructure.
Whether this latest model remains a “minor” step or marks a precursor to something more transformative like R2, one thing is clear: the AI race is no longer a one-sided sprint.