Chinese artificial intelligence startup DeepSeek quietly released an updated version of its R1 inference model early Thursday morning, increasing pressure on U.S. AI leaders such as OpenAI.
The updated model, named R1-0528, was made available on the developer platform Hugging Face, though the company has not issued an official public release or detailed explanation of the model’s capabilities.
Benchmark Rankings Show Slight Gains
Despite the lack of documentation, benchmark results offer a glimpse into R1-0528’s performance. According to LiveCodeBench, a widely tracked leaderboard created by researchers at UC Berkeley, MIT, and Cornell, the new DeepSeek model ranks just below OpenAI’s o4 mini and o3 models in code generation. However, it outperforms xAI’s Grok 3 mini and Alibaba’s Qwen 3.
DeepSeek Calls It a “Small-Scale Upgrade”
Bloomberg first reported the update on Wednesday. A DeepSeek representative confirmed in a WeChat group that the company had completed what it described as a “small-scale test upgrade”, and that users were free to begin experimenting with it.
DeepSeek Gained Early Attention with R1
DeepSeek first gained attention in January, when it released the original R1 model. The model matched or exceeded performance benchmarks of leading U.S. models—while operating at a much lower cost. Its success challenged the idea that U.S. export restrictions and limited access to computing power were preventing Chinese AI development.
The launch of R1 also rattled tech stocks outside of China and shook assumptions that AI progress depends solely on scale and infrastructure.
Rivals Respond with Price Cuts and Lighter Models
Since then, other major Chinese firms—including Alibaba and Tencent—have released models they claim surpass R1. In response, U.S. firms have adjusted their strategies. Google’s Gemini introduced cheaper pricing tiers, and OpenAI launched the o3 mini, a more efficient model with lower hardware requirements.
R2 Expected Soon
The AI community now awaits the official release of R2, the planned successor to R1. According to a March report by Reuters, R2 was initially scheduled to launch in May. In the meantime, DeepSeek did release a V3 upgrade to its large language model earlier this year.
