DeepSeek quietly releases new AI version with high reasoning ability

(Dan Tri) - The global artificial intelligence (AI) race has once again heated up as DeepSeek, an ambitious Chinese AI startup, has quietly released an update to its R1 inference model.

The new version, R1-0528, of DeepSeek has impressive programming capabilities (Photo: Reuters).

The new AI version is called R1-0528 on the Hugging Face platform. This "silent" move is expected to continue to increase competitive pressure on heavyweights from the United States such as OpenAI, in the context of the increasingly fierce global AI race.

Although DeepSeek has not made any official announcement or detailed description of this new "brainchild", experts have quickly made initial assessments.

According to the LiveCodeBench ranking – a reputable metric developed by researchers from UC Berkeley, MIT and Cornell – DeepSeek's updated R1 model shows impressive code generation (programming) capabilities.

It was just slightly behind OpenAI's o4 and o3 mini models, but surpassed xAI's Grok 3 mini (Elon Musk's company) and Alibaba's Qwen 3.

Previously, Bloomberg news agency reported that a DeepSeek representative shared in a WeChat group that this was just a "small experimental upgrade" and encouraged users to start experiencing it.

Earlier this year, DeepSeek made headlines when it challenged the notion that US export controls were stifling China's AI development.

The company has launched AI models that perform on par with, or even outperform, leading US models, at a fraction of the cost.

The launch of R1 in January caused a stir, sending tech stocks outside China plunging and raising questions about whether AI really requires the massive computing power and investment it once thought it did.

Since that impressive debut, the AI ​​race in China has become increasingly fierce with the participation of tech giants such as Alibaba and Tencent, which have also quickly announced their own models, which are said to be more capable than DeepSeek.

In that context, American competitors are not sitting still. Google has introduced discounted Gemini access levels, while OpenAI has also reduced prices and released the o3 Mini model, a more compact version that consumes less computational resources.

Currently, the technology world is still expecting DeepSeek to soon launch R2, the expected successor to R1, which Reuters initially reported would be released in May. In addition, DeepSeek also managed to upgrade its large language model V3 last March.

Clearly, the global AI war is entering an accelerated phase with continuous innovations and relentless competition from major players, promising to bring many breakthroughs in the future.

Source: https://dantri.com.vn/cong-nghe/deepseek-am-tham-tung-phien-ban-ai-moi-co-kha-nang-suy-luan-cao-20250529093858819.htm