Baidu’s ERNIE 5 AI model rises in the rankings: a math whiz that beats OpenAI’s GPT 5.1

Chinese tech giant Baidu just updated its state-of-the-art AI model – and it’s apparently pretty good.

The company’s ERNIE-5.0-0110 scored 1,460 points on LMArena’s Text leaderboard, landing at No. 8 globally and becoming the only Chinese model to reach the platform’s top 10. It ranked higher than OpenAI’s GPT-5.1-High and Google’s Gemini-2.5-Pro. ERNIE v4.0 was released in 2023.

ERNIE 5.0 also claimed the #2 spot globally for mathematical reasoning, behind only the unreleased GPT-5.2-High. That a Chinese model outperforms virtually every publicly available Western system on complex logic tasks represents a significant shift in the AI ​​capability gap – or more accurately, closing it.

Baidu’s announcement on X highlighted that ERNIE 5.0 has officially left the preview stage. The model also demonstrated competitive performance in creative writing, instructional following, and coding, placing it in the top 10 in multiple career categories, including science, business and finance, and healthcare.

The technical architecture behind these results follows China’s focus on efficiency rather than pure energy. ERNIE 5.0 uses a Mixture-of-Experts architecture with roughly 2 trillion parameters, activating only a small fraction of experts (only 3% to be more precise) per inference.

Like other sparse frontier models, this design reduces computing power per query compared to dense systems, although it introduces additional technical complexity. Baidu also said that ERNIE 5.0 is natively trained on text, images, audio and video, rather than retrofitting multimodal modules onto a language-only core. The company said this makes Ernie a native “omni-modal” AI model.

The comeback is remarkable considering how far Baidu had fallen. By November, the preview version of ERNIE 5.0 had fallen to 24th place on LMArena. The domestic AI market was dominated by faster competitors: ByteDance’s Doubao had 101 million monthly active users, while DeepSeek’s cost-efficient models sparked a price war that forced Baidu to completely abandon its paid subscription model last April.

But the consumer woes haven’t slowed Baidu’s entrepreneurial momentum. ERNIE now operates smart city command centers across China, serving all systemically important Chinese banks and processing 16.5 billion API calls daily. This B2B stronghold has insulated Baidu from cutthroat consumer competition while funding its continued model development.

Besides the sheer power of the LLM, the user interface for the ERNIE chatbot is quite user-friendly and packed with pre-customized features that allow users to get more customized results based on their needs. For example, instead of making users deal with complex prompts for different tasks, the chatbot offers separate sections for writing, reading, image editing, and general use. It’s the same core model, but the variations in the system prompts and customizations make every task easier to perform.

Interestingly enough, the most recent version does not include web search, so it is purely offline. However, users can switch to the previous Ernie 4.5 and get updated information while web search is enabled.

Reactions have been mixed, with some users still waiting for a more detailed breakdown of the model’s performance and benchmarks, which the company promised to release soon. It is also important to consider that while ERNIE 5.0 matches GPT-5 and Gemini 2.5 on specific benchmarks, many Western labs have already moved to GPT-5.2 or Gemini 3, making this more of a catch-up than a leap forward.

Users can test the model for free at ernie.baidu.com.