en.Wedoany.com Reported - On May 9, 2026, Baidu officially released its next-generation foundation model, ERNIE 5.1. The model employs "Multi-Dimensional Elastic Pre-training" technology, compressing total parameters to about one-third of ERNIE 5.0 and active parameters to about one-half. Its pre-training cost is only about 6% of comparable industry models. It has already topped the domestic chart and ranked fourth globally on the LMArena search leaderboard, and is simultaneously available on Baidu Qianfan Model Square and the ERNIE Bot official website.
Unlike the mainstream industry path of purely pursuing parameter scale, ERNIE 5.1's technological breakthrough focuses on structural innovation in training efficiency. When ERNIE 5.0 was released, Baidu's R&D team first proposed the "Once-for-All" elastic training framework. Its core logic involves simultaneously optimizing numerous sub-models with varying parameters through a dynamic sampling mechanism during a single pre-training process, constructing a "sub-model matrix" covering multiple parameter scales and computational costs. ERNIE 5.1 extracts the optimal sub-network architecture from this sub-model matrix, fully inheriting the knowledge base of ERNIE 5.0 while achieving leapfrog optimization in both parameter efficiency and training cost.
From a technical detail perspective, the elastic training framework achieves elastic compression and expansion across three dimensions. In terms of elastic depth, it randomly skips some Transformer layers during training, allowing sub-models of different depths to share weights and adaptively learn the balance between deep and shallow representations. For elastic width, it dynamically masks some experts in the MoE layer, forcing the remaining experts to handle more diverse tasks and improving expert utilization efficiency. Regarding elastic sparsity, it flexibly adjusts the number of activated experts through a variable Top-k routing mechanism—activating fewer experts reduces inference costs, while activating more experts enhances model capability, achieving a dynamic balance between inference overhead and performance.
Multiple authoritative benchmark tests have verified the performance level of ERNIE 5.1. In terms of agent capability, ERNIE 5.1 outperforms DeepSeek-V4-Pro on the τ³-bench and SpreadsheetBench-Verified evaluation tasks, with its agentic ability approaching leading international closed-source models. For reasoning ability, it scored 99.6 on the AIME26 math competition evaluation (using tools), second only to Gemini 3.1 Pro. In creative writing capability, internal evaluations show it is approaching Gemini 3.1 Pro. Regarding world knowledge and knowledge understanding, its performance on GPQA and MMLU-Pro evaluations is close to leading closed-source models.
To drive the evolution of large models towards autonomous decision-making agents, Baidu has simultaneously built foundational technology for decoupled fully-asynchronous reinforcement learning, specifically addressing global optimization challenges caused by training-inference discrepancies, low resource utilization, and long-tail effects. Through large-scale agent post-training and environment-expert-fusion full-chain collaboration strategies, the model maintains stable performance when handling complex long-tail tasks. In terms of search capability, ERNIE 5.1 can rapidly retrieve, integrate, and generate information from multi-source data, producing answers with stronger consistency and higher reliability, offering significant practical value in complex business scenarios such as content creation, intelligent assistants, enterprise knowledge management, and agent applications.
Prior to this, the ERNIE 5.0 series had repeatedly topped the LMArena text and visual understanding leaderboards, firmly establishing itself in the first tier of domestic models. On April 30, the ERNIE 5.1 Preview version topped the LMArena text leaderboard domestically with a score of 1476, surpassing mainstream models like GPT-5.5 and DeepSeek-V4-Pro, and was the only domestic model to rank in the top fifteen. The Create 2026 Baidu AI Developer Conference will be held from May 13 to 14 at the China National Convention Center Phase II in Beijing, where Baidu will announce the latest advancements in AI technology breakthroughs and industrial implementation centered around the ERNIE large model.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
