China's AI Models Rise as Innovation Accelerates

China's artificial intelligence ecosystem is gaining rapid global traction, with domestic large models rising in international rankings and expanding their influence across industries.

The latest assessment released by the prominent evaluation platform LMArena ranks several Chinese models among the world's best. Baidu's ERNIE model has been placed second globally — and first in China — for text capabilities, excelling in creative writing, long-form comprehension, and following instruction. It also ranks among the global leaders in visual understanding.

Zhipu's GLM-4.6 has a leading position in coding capabilities, matching the performance of other top international models.

Since the rise of DeepSeek earlier this year, global interest in China's AI development has intensified. Large models have quickly become a strategic battleground in international competition. Through advances in fundamental research, product deployment, and ecosystem building, Chinese AI developers are entering the global first tier, helping to reshape productivity in sectors across the economy.

"Intelligence itself is the biggest application, and the speed of technological iteration is the only real moat," Baidu founder Li Yanhong said at the 2025 Baidu World Conference. Over the past year, he remarked, the capabilities of large models have expanded well beyond conversational tools, with notable progress in digital humans, coding agents, and systems capable of autonomously searching for "globally optimal solutions" in complex scenarios.

China currently leads the world in the pace of model development. By July, Chinese institutions had released 1,509 large models — 40 percent of the global total. According to Stanford University's 2025 AI Index Report, China's high-performance models have rapidly improved in both quantity and quality, with the performance gap between top Chinese and international systems narrowing significantly by the end of 2024.

At the heart of this progress is continuous technological innovation. Baidu's newly launched ERNIE 5.0, a next-generation natively multimodal model with 2.4 trillion parameters, has a unified architecture capable of processing and generating text, images, audio, and video.

Unlike earlier approaches that combined separate modules for different types of data, Baidu's new design builds a single, integrated framework from the ground up, enabling the model to handle heterogeneous information with greater fluidity and synergy, according to Baidu.

Chinese AI models are also making a strong mark on the global open-source community. A joint report by MIT and Hugging Face, a U.S. company that uses machine learning to develop computation tools for building applications, says Chinese developers accounted for 17.1 percent of worldwide downloads of open-source AI models over the past year, surpassing the U.S.

Baidu's multimodal reasoning model recently topped both the global model trending chart and the global multimodal chart on Hugging Face. "Baidu has led more than 1,000 open-source projects, with over 21,000 community contributors," said Chen Shangyi, chair of Baidu's technology committee.

Competition among large models is increasingly shaped by the strength of foundational hardware and development platforms. Domestic AI chipmakers such as Kunlunxin, Ascend, and Cambricon are rising rapidly, with Baidu's latest Kunlunxin products offering significant improvements in performance and cost.

At the software level, platforms including Baidu's PaddlePaddle, Huawei's MindSpore, and Megvii's MegEngine are helping developers deploy large-model applications more efficiently.

As technological capacity expands, real-world applications are emerging as the primary measure of value. With its vast population and diverse industry scenarios, China has become the world's largest and most dynamic testing ground for large-model deployment.

In the energy sector, for example, Baidu worked with State Grid to build a unified AI innovation platform and deploy an intelligent drone inspection system capable of five million tower inspections annually, reducing manual climbs by 40 percent.

Backed by an end-to-end AI stack — from chips to frameworks, models, and applications — Baidu is partnering with enterprises across education, healthcare, finance, and manufacturing. Its PaddlePaddle-ERNIE ecosystem now supports over 23 million developers and 760,000 companies.

"When AI becomes a native capability," Li said, "intelligence is no longer a cost — but a source of productivity."

Source: Science and Technology Daily