Baidu released ERNIE 5.1, claiming it cost 94% less to train than comparable AI systems, achieved through 'multi-dimensional elastic pre-training' which extracted a sub-network from ERNIE 5.0, reducing total parameters to one-third and active parameters by half. Despite lower training costs, ERNIE 5.1 ranks fourth globally on the LMArena Search leaderboard and first among Chinese models. Its agentic capabilities surpass DeepSeek-V4-Pro. The efficiency mirrors DeepSeek R1's impact on inference costs, but here it's training-side savings. Baidu used a four-stage reinforcement learning system (MOPD) with parallel expert models for code, reasoning, and agentic tasks, then distilled into one model. On GPQA and AIME26, ERNIE 5.1 approaches leading closed-source models. It is now deployed on over 10 platforms in China, including AI roleplay and short drama generation. The model is accessible via ernie.baidu.com and API. Baidu's Create 2026 conference on May 13-14 will showcase applications.

Key facts

• ERNIE 5.1 training cost is 6% of comparable AI models, using multi-dimensional elastic pre-training.
• Ranked 4th globally on LMArena Search leaderboard, 1st among Chinese models.
• Total parameters reduced to one-third of ERNIE 5.0; active parameters halved.
• Four-stage MOPD reinforcement learning with parallel expert models for code, reasoning, and agentic tasks.
• Assessed on AIME26 with 99.6% score using tool-assisted reasoning, trailing only Gemini 3.1 Pro.