Baidu’s ERNIE 5.1 cuts training cost 94%, ranks fourth

Baidu released ERNIE 5.1 last week, a compressed successor to ERNIE 5.0 that reduced estimated training cost by roughly 94% compared with similar large models. The company reported the model scored 1,223 on the LMArena Search leaderboard, placing fourth globally and first among Chinese models.

Engineers reduced ERNIE’s total parameters to about one-third of the predecessor and cut active parameters-the parts used during a conversation-by roughly half. The smaller network retained the parent model’s knowledge without repeating a full pre-training cycle.

Baidu extracted an optimized sub-network from ERNIE 5.0 using a process the company calls multi-dimensional elastic pre-training. Baidu framed the training expense as about 6% of what comparable models typically require.

A four-stage post-training pipeline named MOPD, short for Multi-Teacher On-Policy Distillation, combined specialist expert models trained in parallel for coding, reasoning and agentic tasks into a single model. A final online reinforcement learning stage adjusted open-ended conversation and creative outputs.

On agentic tasks such as filling spreadsheets and autonomous web browsing, ERNIE 5.1 outperformed DeepSeek-V4-Pro in LMArena evaluations. On specialist benchmarks the model approached the performance of leading closed-source western models on GPQA and recorded 99.6% on the 2026 American Invitational Mathematics Examination when using tool-assisted reasoning, trailing only Gemini 3.1 Pro.

Baidu reported ERNIE 5.1 is being deployed across more than 10 creative and agentic platforms in China, including AI roleplay services and short-drama generation tools. The model is available at ernie.baidu.com and through Baidu AI Cloud’s API.

Baidu plans to demonstrate ERNIE applications at its Create 2026 developer conference in Beijing on May 13–14. The company launched Ernie Bot in August 2023 and reported the chatbot reached 100 million users in China by December 2023. ERNIE 5.0, the parent model, was published in January 2026.

The release follows a trend in Chinese AI research toward efficiency in model design and training. In January 2025 the startup DeepSeek released R1, a model that sharply reduced inference cost and prompted reevaluation of compute strategies across the industry.