Chinese AI startup DeepSeek quietly updated its R1 reasoning model early Thursday, intensifying competition with U.S. companies like OpenAI.
The new version, R1-0528, was launched on the developer platform Hugging Face without an official announcement or detailed description. However, the LiveCodeBench leaderboard, a code generation benchmark created by researchers at UC Berkeley, MIT, and Cornell, places DeepSeek’s updated model just behind OpenAI’s o4 mini and o3 reasoning models, while outperforming xAI’s Grok 3 mini and Alibaba’s Qwen 3.
Earlier reports said DeepSeek described this as a “minor trial upgrade” and encouraged users to start testing it.
Earlier this year, DeepSeek surprised the tech world by releasing AI models comparable to or better than top U.S. models at a much lower cost, challenging assumptions that U.S. export controls were limiting China’s AI progress. The launch of R1 in January shook global tech markets and questioned the notion that massive computing power and investments are essential for leading AI development.
Since then, other Chinese tech giants, such as Alibaba and Tencent, have launched models aiming to surpass DeepSeek’s performance. Meanwhile, Google introduced more affordable access to its Gemini AI, and OpenAI reduced prices and launched the more efficient o3 mini model.
DeepSeek is expected to release R2, the next iteration of its reasoning model. Reports earlier this year suggested the release was planned for May. In March, DeepSeek also upgraded its V3 large language model.
