Wenxin 5.1 - Baidu's flagship large language model

What is Wenxin 5.1?

ERNIE 5.1 is Baidu's flagship large-scale language model. Based on ERNIE 5.0's knowledge distillation and Once-for-All elastic training framework, it reduces pre-training costs to 6% of industry-standard models of similar scale. The model excels in agent-based tasks, deep search, mathematical reasoning, and creative writing, ranking first in China and fourth globally on the Arena search leaderboard. Employing separate, fully asynchronous reinforcement learning and multi-stage online policy distillation techniques, the model supports access via the ERNIE official website, Qianfan API, and Xinghe Community, setting a benchmark for cost-effectiveness among domestic large-scale language models.

The main functions of Wenxin 5.1

• Agent's outstanding capabilitiesIn the τ3-bench and SpreadsheetBench-Verified benchmarks, its performance surpasses DeepSeek-V4-Pro and approaches the level of top-tier closed-source models.
• Leading in depth searchWith a score of 1223, it ranked fourth globally and first in China on the Arena search rankings, demonstrating its ability to retrieve complex information and reason about knowledge.
• World Knowledge and Reasoning ExcellenceIts GPQA and MMLU-Pro benchmarks are close to those of top closed-source models, and its AIME26 math competition score of 99.6 is second only to Gemini-3.1 Pro.
• Outstanding creative writing skillsInternal reviews indicate it is close to the Gemini 3.1 Pro, and it ranks first in Text Arena in China, gaining widespread recognition from professional writers and content platforms.
• Ultimate cost-effectivenessBased on multi-dimensional elastic pre-training technology, the pre-training cost is only 6% of that of similar-sized models in the industry, achieving the lowest training overhead for the same level of performance.

Technical Principles of Wenxin 5.1

• Once-for-All Flexible Training FrameworkTraditional methods require pre-training models of different scales separately. Wenxin 5.0 optimizes a large number of sub-models with different parameters simultaneously in a single pre-training process through a dynamic sampling mechanism, and constructs a sub-model matrix covering multiple parameter scales and computational costs. Wenxin 5.1 then extracts the optimal substructure from it and inherits the knowledge.
• Multi-dimensional elastic compressionIn terms of elastic depth, some Transformer layers are randomly skipped to allow sub-models of different depths to share weights; in terms of elastic width, MoE experts are dynamically masked to force the remaining experts to undertake more diverse tasks; in terms of elastic sparsity, the number of activated experts is flexibly adjusted through variable Top-k routing to achieve a dynamic trade-off between inference overhead and performance.
• Separate Fully Asynchronous Reinforcement LearningWith the RL Controller at its core, the control planes of the four subsystems—training, inference, reward, and agent loop—are completely decoupled. Each subsystem is deployed independently and can be scaled up or down. Inference and training form a natural pipeline that can fully overlap and mask each other, laying a highly scalable foundation for long-term asynchronous training.
• FP8 training-push consistency optimizationBased on the PaddlePaddle training and propulsion integrated framework, a unified FP8 low-precision operator library is implemented to minimize the training and propulsion accuracy deviation in reinforcement learning; Rollout Router Replay technology is optimized for the MoE model, and the K3 KL divergence is reduced by 50% through two-stage computation and communication masking and dynamic bit compression.
• Multi-stage OPD training pipelineAfter adopting a unified SFT initialization, expert models in the fields of code, inference, and agent are trained in parallel to avoid mutual interference; then, the capabilities of multiple experts are integrated into a unified parameter space through online policy distillation of token-level reverse KL; finally, general capabilities such as open chat and creative writing are improved through general online RL.

How to use Wenxin 5.1

• Official website chat experienceVisit the Wenxin Yiyan official website to interact directly with the model.
• API call service: Use the Qianfan large model platform (https://console.bce.baidu.com/qianfan/) and modify the model_name to ernie-5.1 to call it.
• Playground ExperienceLog in to Xinghe Community https://aistudio.baidu.com/playground to start the Wenxin 5.1 online experience.

The core advantages of Wenxin 5.1

• Ultimate cost-effectivenessThe pre-training cost is only 6% of that of similar-sized models in the industry, the total parameters are compressed to 1/3 and the activation parameters are compressed to 1/2, achieving the lowest training cost at the same level of performance.
• Search capabilities ranked first in ChinaArena Search ranked fourth globally and first in China with a score of 1223, surpassing models such as GPT-5.5 Search.
• Agent and reasoning abilities are outstanding.τ3-bench surpasses DeepSeek-V4-Pro, AIME26 score of 99.6 is close to Gemini-3.1 Pro, and it has the ability to break down complex tasks and call tools.
• Creative ability recognized by the industryRanked 13th in Text Arena (first in China), he has received praise from professional content platforms and writers such as Phoenix Smart Media and Origin Concept, and is skilled in script logic and emotional control.
• Advanced training architectureThe separate, fully asynchronous RL infrastructure supports independent scaling and heterogeneous resource scheduling, and FP8 training-push consistency ensures the stability of long-term training.

The project address for Wenxin 5.1

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

Comparison of Wenxin 5.1 with similar competing products

Application Scenarios of Wenxin 5.1

• Deep Search and Knowledge ResearchSuitable for scenarios requiring high accuracy, such as complex information retrieval, academic literature review, and cross-disciplinary knowledge organization.
• Agent automation tasksIt supports the automated execution of technical solutions such as code generation, table processing, multi-step workflow design, and requirements analysis.
• Creative content productionIt serves content industries that require narrative logic and style matching, such as scriptwriting, novel writing, professional copywriting, and advertising creativity.
• Mathematical and Logical ReasoningIt can handle high-intensity reasoning scenarios such as competition-level mathematical problems, technical solution derivation, and logical verification.
• Enterprise-level AI integrationBy connecting to enterprise business systems through the Qianfan platform API, B-end applications such as customer service, knowledge base, and intelligent office can be built.