ERNIE 5.1 - Baidu's flagship large language model

What is ERNIE 5.1

ERNIE 5.1 is Baidu's flagship large language model. Built on ERNIE 5.0 knowledge distillation and the Once-for-All elastic training framework, it reduces pre-training cost to 6% of industry models of the same scale. The model excels in Agent tasks, deep search, mathematical reasoning, and creative writing, ranking first in China and fourth globally on the Arena search leaderboard. The model adopts decoupled fully asynchronous reinforcement learning and multi-stage online policy distillation technologies, and supports invocation throughERNIE Botofficial website, Qianfan API, and Xinghe Community, positioning itself as a cost-effectiveness benchmark for domestic large models.

Main features of ERNIE 5.1

• Outstanding Agent capabilities: In the τ3-bench and SpreadsheetBench-Verified evaluations, its performance surpasses DeepSeek-V4-Pro, approaching the level of top closed-source models.
• Leading deep search: With a score of 1223, it won fourth place globally and first place in China on the Arena search leaderboard, validating its complex information retrieval and knowledge reasoning capabilities.
• Excellent world knowledge and reasoning: On GPQA and MMLU-Pro evaluations it approaches top closed-source models, and its AIME26 math competition score of 99.6 is second only to Gemini-3.1 Pro.
• Outstanding creative writing ability: In internal evaluations it approaches Gemini 3.1 Pro, ranks first in China on Text Arena, and has been widely recognized by professional writers and content platforms.
• Extreme cost-effectiveness: Based on multi-dimensional elastic pre-training technology, the pre-training cost is only 6% of industry models of the same scale, achieving the lowest training overhead at the same level of performance.

Technical principles of ERNIE 5.1

• Once-for-All elastic training framework: Traditional methods require separate pre-training for models of different scales. ERNIE 5.0, in a single pre-training run, uses a dynamic sampling mechanism to simultaneously optimize a large number of sub-models with different parameters, building a sub-model matrix covering multiple parameter scales and computational costs. ERNIE 5.1 extracts the optimal substructure from it and inherits the knowledge.
• Multi-dimensional elastic compression: In elastic depth, some Transformer layers are randomly skipped so sub-models of different depths share weights; in elastic width, MoE experts are dynamically masked to force the remaining experts to take on more diverse tasks; in elastic sparsity, variable Top-k routing flexibly adjusts the number of activated experts, achieving a dynamic trade-off between inference cost and performance.
• Decoupled fully asynchronous reinforcement learning: With the RL Controller as the core, the control planes of the four major subsystems—training, inference, reward, and Agent Loop—are fully decoupled. Each subsystem is independently deployed and scaled, and inference and training form a natural pipeline that can be fully overlapped and masked, laying a highly scalable foundation for long-horizon asynchronous training.
• FP8 training-inference consistency optimization: Based on the PaddlePaddle integrated training-inference framework, it implements a unified FP8 low-precision operator library, minimizing training-inference precision deviation in reinforcement learning; for MoE models, it optimizes Rollout Router Replay technology, and through two-stage computation-communication masking and dynamic bit compression, reduces K3 KL divergence by 50%.
• Multi-stage OPD training pipeline: After unified SFT initialization, domain expert models in code, reasoning, Agent, and other fields are trained in parallel to avoid mutual interference; then, through token-level reverse KL online policy distillation, the capabilities of multiple experts are fused into a unified parameter space; finally, general online RL improves general capabilities such as open-ended chat and creative writing.

How to use ERNIE 5.1

• Official website chat experience: VisitERNIE BotOfficial website: Chat directly with the model.
• API Calling Service: Via the Qianfan Large Model Platform https://console.bce.baidu.com/qianfan/ modify the model_name to ernie-5.1 to make calls.
• Playground Experience: Log in to the Galaxy Community https://aistudio.baidu.com/playground to start the Wenxin 5.1 online experience.

Core Advantages of Wenxin 5.1

• Ultimate Cost-Effectiveness: The pretraining cost is only 6% of industry models of the same scale, with total parameters compressed to 1/3 and activated parameters compressed to 1/2, achieving the lowest training overhead at the same level of performance.
• No. 1 in China for Search Capability: Ranked fourth globally and first in China on the Arena search leaderboard with 1223 points, surpassing models such as GPT-5.5 Search.
• Outstanding Agent and Reasoning Capabilities: τ3-bench surpasses DeepSeek-V4-Pro, and the AIME26 score of 99.6 approaches Gemini-3.1 Pro, with capabilities for decomposing complex tasks and tool calling.
• Creative Ability Recognized by the Industry: Ranked 13th in Text Arena (first in China), praised by professional content platforms and writers such as Phoenix Zhimei and Yuandian Gouxiang, and skilled at script logic and emotional control.
• Advanced Training Architecture: The separated fully asynchronous RL infrastructure supports independent scaling and heterogeneous resource scheduling, while FP8 training-inference consistency ensures stability in long-horizon training.

Project Address of Wenxin 5.1

• Project Official Website：https://yiyan.baidu.com/blog/posts/ernie-5.1-0508-release/

Comparison of Wenxin 5.1 with Similar Competitors

Application Scenarios of Wenxin 5.1

• Deep Search and Knowledge Research: Suitable for scenarios requiring high accuracy, such as complex information retrieval, academic literature review, and cross-domain knowledge organization.
• Agent Automation Tasks: Supports automated execution of technical solutions such as code generation, table processing, multi-step workflow design, and requirements analysis.
• Creative Content Production: Serves the content industry that requires narrative logic and style adaptation, such as scriptwriting, novel writing, professional copywriting, and advertising creativity.
• Mathematics and Logical Reasoning: Can handle high-intensity reasoning scenarios such as competition-level mathematics problems, technical solution derivation, and logical verification.
• Enterprise-Level AI Integration: Connects to enterprise business systems through the Qianfan Platform API to build B-end applications such as customer service, knowledge bases, and intelligent office work.