Baidu, China’s search giant trading as BIDU on the Nasdaq, has released ERNIE 5.1, claiming its training cost is 94% lower than comparable frontier AI systems. The model employs a novel 'multi-dimensional elastic pre-training' approach, extracting an optimized sub-network from its predecessor ERNIE 5.0. Total parameters were compressed to about one-third, while active parameters were halved, allowing the model to inherit a vast knowledge base without the full expense.
On the LMArena Search Arena, ERNIE 5.1 scored 1,223, placing fourth globally and first among all Chinese models. Its agentic capabilities-handling multi-step tasks like autonomous web browsing-surpassed DeepSeek-V4-Pro, the previous Chinese leader. On the AIME26 math competition, it achieved a 99.6% score using tool-assisted reasoning, trailing only Gemini 3.1 Pro.
The post-training pipeline uses a four-stage reinforcement learning system called MOPD. Specialist expert models for code, reasoning, and agentic tasks were trained in parallel and distilled into a single unified model, avoiding performance trade-offs.
ERNIE 5.1 is already being deployed across more than 10 creative and agentic platforms in China, including AI roleplay and short drama generation tools. Baidu’s annual Create 2026 developer conference on May 13-14 in Beijing will showcase further industrial applications.