Overview

• DeepSeek quietly launched the R1-0528 trial upgrade on Hugging Face without a formal public announcement, describing the changes as a minor update in a company WeChat group.
• LiveCodeBench benchmarks show R1-0528 trailing only OpenAI’s o4-mini and o3 reasoning models in code generation while outperforming xAI’s Grok 3 mini and Alibaba’s Qwen 3.
• This is the first revision of the January-released R1 reasoning model, which initially disrupted global tech stocks by proving cost-efficient AI development viable.
• DeepSeek developed the original R1 at a reported cost of about $6 million, undercutting Western rivals that have spent billions on comparable AI model training.
• Despite its technical advances, DeepSeek faces oversight under China’s internet regulations enforcing ‘core socialist values’ and usage restrictions by some governments and corporations over data security concerns.