Baidu ERNIE-5.1-Preview: Efficient MoE Leader Arrives

Introduction

Baidu has officially unveiled ERNIE-5.1-Preview, a significant leap in large language model efficiency and performance released on May 9, 2026. This model addresses the industry's growing need for cost-effective inference without sacrificing intelligence. Unlike previous iterations that relied on brute-force scaling, ERNIE-5.1-Preview introduces a highly optimized architecture designed for enterprise-grade applications.

The release marks a strategic pivot for Baidu, positioning their model to compete directly with global leaders like DeepSeek and OpenAI's latest offerings. By focusing on parameter efficiency and specialized post-training, Baidu aims to democratize access to high-quality AI capabilities for developers and businesses operating within the Chinese market and globally.

• Release Date: 2026-05-09
• Provider: Baidu
• Category: Language Model

Key Features & Architecture

The core innovation of ERNIE-5.1-Preview lies in its Mixture of Experts (MoE) architecture. Baidu has successfully compressed the total parameters to approximately one-third of the previous ERNIE-5.0 model, while active parameters are reduced to roughly half. This drastic reduction is achieved without compromising the model's ability to process complex tasks, demonstrating advanced sparse training techniques.

Technically, the model is powered by a decoupled fully-asynchronous reinforcement learning infrastructure built on Baidu's PaddlePaddle framework. It utilizes a scaled agentic post-training pipeline featuring Multi-Teacher On-Policy Distillation (MOPD). This four-stage pipeline includes Supervised Fine-Tuning (SFT), Domain Expert Model Training, On-Policy Distillation, and General Online RL, ensuring robust knowledge retention.

• Total Parameters: ~1/3 of ERNIE-5.0
• Active Parameters: ~1/2 of ERNIE-5.0
• Training Cost: ~6% of comparable models
• Framework: PaddlePaddle

Performance & Benchmarks

ERNIE-5.1-Preview has demonstrated exceptional performance across specialized leaderboards, topping the LMArena Search leaderboard as #4 globally and #1 among Chinese models with a score of 1,223. Its specialized capabilities are particularly noteworthy, securing the #1 global rank in Legal & Government and #4 in Business, Management & Financial Ops.

In terms of reasoning and coding, the model achieves a score of 99.6 on AIME26 with tool use, ranking second only to Gemini 3.1 Pro. It also surpasses DeepSeek-V4-Pro on critical agent evaluation tasks like tau-cubed-bench and SpreadsheetBench-Verified. This performance is achieved at a fraction of the pre-training cost, highlighting the efficiency of the underlying architecture.