China’s DeepSeek closes in on US rival OpenAI, surpasses Alibaba with upgraded model
Benchmark results cited by DeepSeek show that R1-0528’s performance surpasses Alibaba’s Qwen3 and matches models from OpenAI, Google
In a statement released late on Thursday, DeepSeek highlighted improvements in the new model’s reasoning and creative writing capabilities, making it more adept at crafting argumentative essays, fiction and prose in styles that closely mimic human authors. Coding capabilities have also been enhanced.
The company said the latest version achieved a 50 per cent reduction in “hallucinations” – instances where AI generates misleading information with little factual basis.
These upgrades were achieved by investing additional computing resources in the post-training stage, when developers make final adjustments and enhancements to the model after the main training process, the company said. Post-training usually focuses on boosting efficiency and enhancing content safety and accuracy.
“The updated R1 model excelled among domestic AI models in a range of benchmark tests, including maths, coding and general logic, and matched up to global top models such as [OpenAI’s] O3 and [Google’s] Gemini2.5-Pro,” DeepSeek said.
