Baidu’s New AI Is Already Beating Top Models And Dropped By 94% Productivity

Baidu, China’s equivalent of Google, officially released ERNIE 5.1 last weekend, and the headline is hard to ignore: Costs about 94% less to train than comparable AI systems at the same level.

To that end, training a borderline AI model usually costs a few bucks millions (if not billions) of dollars in calculations. Baidu, which controls 76% of China’s search market and trades on Nasdaq as BIDU, says it has left the same class in the past two decades.

The trick is called “multi-dimensional elastic pre-training.” Instead of building ERNIE 5.1 ​​from scratch, Baidu removed the optimized network from its ERNIE 5.0 design – released in January 2026 – and compressed it. The total number has dropped to one-third of the original. What worked (what actually worked during the discussion) was cut in half. The result is a lean model that took the knowledge of its great parent without repeating the entire training bill.

On LMArena Search Arena — a board where real users compare AI models for online searches, powered by human preferences—ERNIE 5.1 ​​scored 1,223it has ranked fourth in the world and first among all Chinese brands. Its technical capabilities – the way it handles multi-tasking tasks such as filling out spreadsheets or browsing the web on its own – outperformed DeepSeek-V4-Pro, an older Chinese setter.

ERNIE may sound like a strange name among western viewers, but it is actually a great example in China. Baidu launched Ernie Bot in August 2023, and chatbot reached 100 million users in China by December of that year—faster than most of its Chinese competitors, though slower than ChatGPT’s world record of two months reaching the same level.

The look is similar to what DeepSeek did to the AI ​​industry in January 2025. When the Chinese startup released the R1 – a model that similar to OpenAI’s o1 at 98% low query rate-caused a $600 billion loss in Nvidia’s market value and forced every major AI lab to reconsider whether throwing computers at the problem is the only viable option. ERNIE 5.1 ​​is a different kind of success story – on the educational side rather than the simulation – but the message is the same: Chinese labs find ways to do more with less.

Post-academic pipelines are also important to note. Baidu has developed a four-part reinforcement learning system they call MOPD (Multi-Teacher On-Policy Distillation). Instead of trying to teach every skill at the same time – which tends to introduce “styles” in which, for example, improving math skills to create writing skills – Baidu taught experts similar types of coding, reasoning, and technical work, and then melted them all into one coherent model. The final part of the online motivational study consisted of open discussion and elicitation, keeping in mind what the distillation method could not do well.

In theory it should mean that all skills are developed according to skill, without putting it in front of the other

On GPQA (Graduate-Level Google-Proof Q&A, a benchmark to test whether an artist can answer technical scientific questions that cannot be mentioned by Google), ERNIE 5.1 ​​approaches the performance of leading western closed models. On AIME26 – the American Invitational Mathematics Examination revised in 2026, which tries to solve problems in competitions – the brand scored 99.6% using hardware-assisted reasoning, trailing only Gemini 3.1 Pro.

Baidu says ERNIE 5.1 ​​has already started rolling out to more than 10 platforms in China, including AI platforms and short-form gaming tools. The example can be found at ernie.baidu.com and through an API on Baidu’s AI Cloud platform.

Baidu will host its annual Make 2026 conference on May 13-14 in Beijing, where it plans to showcase the latest ERNIE software. That event will be the next indication of how the company wants to push the model into the business markets and around the world.