DeepSeek Unveils Major Upgrade to Match Rivals Like ChatGPT
By Liu Peilin and Denise Jia
Chinese AI startup DeepSeek has released a major upgrade to its flagship language model, DeepSeek R1 — marking the first significant update in five months — claiming its overall capabilities now approach its major rivals.
The company quietly launched the new version, DeepSeek-R1-0528, on open-source platform Hugging Face on Wednesday morning, followed by a formal announcement later in the day detailing its capabilities and availability.
Users can now access the upgraded model via DeepSeek’s official website, app, or WeChat mini program. According to the company, simply enabling the “Deep Thinking” feature in the chat interface allows users to experience the latest iteration. The model’s application programming interface (API) has also been updated.
While still based on the DeepSeek V3 base model released in December, the new version benefits from substantial increases in post-training compute power. This has reportedly led to a marked improvement in reasoning ability and depth of thought, bringing it closer to top-tier models such as OpenAI’s GPT-4o and Google’s Gemini 2.5 Pro.
Benchmark tests underscore the improvements. On the 2025 American Invitational Mathematics Examination, the upgraded model achieved an accuracy rate of 87.5%, a significant leap from the previous version’s 70%. The model now consumes roughly 23,000 tokens per question, nearly double the earlier version, suggesting deeper and nuanced reasoning processes.
DeepSeek also claims the R1-0528 model reduces hallucinations — AI-generated inaccuracies —by 45% to 50% in tasks such as rewriting, summarization and reading comprehension. The model has been fine-tuned to handle a variety of writing styles, including argumentative essays, fiction and prose, producing longer, more coherent writing in a style better suited to human tastes.
In tandem with the R1 upgrade, DeepSeek released a smaller model, Qwen3-8B, trained on the updated version. Despite its smaller size, the company says Qwen3-8B matches the performance of models with more than 200 billion parameters, such as Qwen3-235B.
Tech giant Tencent has already integrated DeepSeek’s latest model into its personal knowledge management tool, “ima.” The company announced the integration just hours after the open-source release, marking one of the fastest deployments to date. Tencent says ima now supports both its proprietary Hunyuan model and DeepSeek simultaneously, with weekly updates planned.
Since launching its first reasoning model, R1, in December 2023, DeepSeek has experienced a meteoric rise. The model went viral in January, drawing global attention to China’s AI industry. In February, DeepSeek founder Liang Wenfeng was among several top entrepreneurs — including Huawei’s Ren Zhengfei and Alibaba’s Jack Ma — invited to a high-level roundtable with Chinese President Xi Jinping.
The debut of R1 led to a surge in AI development across China’s tech ecosystem, fueling fierce competition in compute infrastructure, model development and AI applications. Within two weeks, infrastructure providers rolled out products optimized for DeepSeek, while rival model developers rushed to open source their work. Chinese firms including Alibaba, Baidu, Tencent and ByteDance have all since released their own models to compete with DeepSeek R1.
Contact reporter Denise Jia (huijuanjia@caixin.com)
caixinglobal.com is the English-language online news portal of Chinese financial and business news media group Caixin. Global Neighbours is authorized to reprint this article.
Image: Artistic Visions – stock.adobe.com