Podcast Episode
Baidu's ERNIE 5.0 Becomes First Chinese AI Model to Crack Global Top 10
January 17, 2026
Audio archived. Episodes older than 60 days are removed to save server storage. Story details remain below.
Baidu released its latest AI model this week, with ERNIE 5.0-0110 achieving a score of 1,460 points on the LMArena Text benchmark and securing eighth place globally. This makes it the only Chinese model to crack the top 10 on the influential ranking platform. The model outperformed OpenAI's GPT 5.1-High and Google's Gemini 2.5-Pro, while achieving second place globally in mathematical reasoning, trailing only GPT 5.2-High.
The ranking, released January 15 by LMArena, comes as Google DeepMind CEO Demis Hassabis told CNBC's podcast The Tech Download that Chinese AI models are now just a matter of months behind their Western counterparts. This statement challenges previous assumptions about a wider technology gap between Chinese and Western AI development.
The model demonstrated competitive capabilities across multiple dimensions, including creative writing, instruction following, and programming tasks. According to Baidu's official announcement, the removal of the Preview label indicates that ERNIE 5.0-0110 has completed its testing phase and entered formal deployment. Baidu stated that achieving the top position among Chinese models upon entering the leaderboard as a formal version demonstrates the model's continuous breakthroughs in general text capabilities.
The achievement in mathematical reasoning is particularly notable, as it represents a domain where AI models have historically struggled with complex logical operations. ERNIE 5.0's second place global ranking in this category signals significant progress in areas requiring deep reasoning capabilities.
However, Hassabis questioned whether Chinese firms can achieve breakthrough innovations rather than simply matching existing frontiers. Can they create something new that surpasses the current boundaries? he said on CNBC. That has not been evidenced yet. This distinction between replication and innovation highlights a key question about whether Chinese AI companies can move from following established approaches to pioneering entirely new paradigms.
The DeepMind CEO noted that inventing something new is approximately 100 times harder than copying existing technology. He attributed the lack of frontier breakthroughs to mentality rather than technical restrictions, suggesting that the challenge is more about fostering a culture of radical innovation than overcoming technological limitations.
Meanwhile, Chinese corporations led by Huawei Technologies and Baidu have positioned China as the global leader in patent filings for physical AI used in humanoid robots and vehicles. This suggests that Chinese companies are making strategic investments across multiple AI domains, not just large language models.
DeepSeek has also signaled ambitious plans for 2026, publishing a technical paper co-authored by founder Liang Wenfeng that proposes rethinking the fundamental architecture used to train foundational AI models. A new reasoning model is expected to be released, which observers believe may bring another major shift in competitive dynamics.
This infrastructure gap represents a serious constraint on Chinese AI ambitions. While Chinese companies have demonstrated impressive efficiency in model development, the sheer scale of computational resources available to American companies provides a substantial advantage in training increasingly large and capable models.
Alibaba has pledged to invest at least 380 billion yuan, approximately 52.4 billion dollars, in its AI and cloud computing infrastructure over the next 3 years. This exceeds what the company invested in those areas over the previous decade, signaling the strategic importance Chinese tech giants are placing on closing the infrastructure gap.
The rapid progress of ERNIE 5.0 and other Chinese models demonstrates that the global AI race is far more competitive than previously assumed. While questions remain about whether Chinese firms can pioneer breakthrough innovations, their ability to rapidly close the gap with Western models has reshaped perceptions of the technological balance between East and West.
Technical Architecture and Capabilities
ERNIE 5.0 is built on a Mixture-of-Experts architecture featuring over 2 trillion parameters, placing it among the largest publicly disclosed AI models. Despite its massive scale, the model activates less than 3 percent of its parameters during each inference, reducing computational costs while maintaining performance. This design allows the model to match the capabilities of larger systems while requiring significantly less processing power for each query.The model demonstrated competitive capabilities across multiple dimensions, including creative writing, instruction following, and programming tasks. According to Baidu's official announcement, the removal of the Preview label indicates that ERNIE 5.0-0110 has completed its testing phase and entered formal deployment. Baidu stated that achieving the top position among Chinese models upon entering the leaderboard as a formal version demonstrates the model's continuous breakthroughs in general text capabilities.
The achievement in mathematical reasoning is particularly notable, as it represents a domain where AI models have historically struggled with complex logical operations. ERNIE 5.0's second place global ranking in this category signals significant progress in areas requiring deep reasoning capabilities.
Narrowing the AI Gap
Hassabis's assessment that China is much closer to the cutting edge models in the United States or the West than we expected 1 to 2 years ago represents a notable shift in perspective from a leader at one of the world's foremost AI research labs. He cited models from DeepSeek, Alibaba, and ZhipuAI as examples of China's rapid progress in developing competitive AI systems.However, Hassabis questioned whether Chinese firms can achieve breakthrough innovations rather than simply matching existing frontiers. Can they create something new that surpasses the current boundaries? he said on CNBC. That has not been evidenced yet. This distinction between replication and innovation highlights a key question about whether Chinese AI companies can move from following established approaches to pioneering entirely new paradigms.
The DeepMind CEO noted that inventing something new is approximately 100 times harder than copying existing technology. He attributed the lack of frontier breakthroughs to mentality rather than technical restrictions, suggesting that the challenge is more about fostering a culture of radical innovation than overcoming technological limitations.
Broader Chinese AI Momentum
The ERNIE 5.0 release coincides with broader gains for Chinese AI companies. Alibaba's Qwen AI app surpassed 100 million monthly active users as of January after recording 149 percent month over month growth in November, making it the world's fastest growing AI application during that period. This rapid user adoption demonstrates strong domestic demand for AI services in China.Meanwhile, Chinese corporations led by Huawei Technologies and Baidu have positioned China as the global leader in patent filings for physical AI used in humanoid robots and vehicles. This suggests that Chinese companies are making strategic investments across multiple AI domains, not just large language models.
DeepSeek has also signaled ambitious plans for 2026, publishing a technical paper co-authored by founder Liang Wenfeng that proposes rethinking the fundamental architecture used to train foundational AI models. A new reasoning model is expected to be released, which observers believe may bring another major shift in competitive dynamics.
Challenges Ahead
Yet significant challenges remain for Chinese AI development. Lin Junyang, technical lead for Alibaba's Qwen team, said at a recent Beijing conference that the likelihood of a Chinese company surpassing U.S. tech giants in AI within 3 to 5 years is less than 20 percent. He cited U.S. computing infrastructure that is one to two orders of magnitude larger than China's as a fundamental bottleneck.This infrastructure gap represents a serious constraint on Chinese AI ambitions. While Chinese companies have demonstrated impressive efficiency in model development, the sheer scale of computational resources available to American companies provides a substantial advantage in training increasingly large and capable models.
Alibaba has pledged to invest at least 380 billion yuan, approximately 52.4 billion dollars, in its AI and cloud computing infrastructure over the next 3 years. This exceeds what the company invested in those areas over the previous decade, signaling the strategic importance Chinese tech giants are placing on closing the infrastructure gap.
The rapid progress of ERNIE 5.0 and other Chinese models demonstrates that the global AI race is far more competitive than previously assumed. While questions remain about whether Chinese firms can pioneer breakthrough innovations, their ability to rapidly close the gap with Western models has reshaped perceptions of the technological balance between East and West.
Published January 17, 2026 at 10:14pm
