This glossary page defines Baidu ERNIE in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.

Eyebrow | AI Foundation Model Family

Baidu ERNIE: China's Large Language Model Family

First released in 2019 to outperform Google's BERT on Chinese-language tasks, ERNIE has grown into a multi-generation model family reaching 2.4 trillion parameters by 2025.

Published: 2026-07-04

Baidu ERNIE is a Product that provides a family of pretrained language and multimodal foundation models for researchers, developers, and enterprises building Chinese-language and multimodal AI applications. Baidu ERNIE belongs to the large language model and foundation-model segment. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems.

Baidu ERNIE: Entity Summary

Entity: Baidu ERNIE (百度文心大模型; full name: Enhanced Representation through Knowledge Integration)
Type: Product (family of large language and multimodal foundation models)
Founded / Launched: The original ERNIE 1.0 research paper was published on arXiv on March 19, 2019, by a Baidu Research team
Founder / Creator: Baidu, Inc., through its Baidu Research division; the founding paper lists authors Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu
Current Owner / Operator: Baidu, Inc.
Headquarters: Beijing, China
Official Website: https://ernie.baidu.com/
Primary Language: Chinese-focused, with multilingual and multimodal capability in later versions
Status: Active
Synonyms / Aliases: 文心大模型 (Wenxin Large Model); ERNIE; Enhanced Representation through Knowledge Integration
Category: Large language model / foundation model family

Baidu ERNIE: Core Facts

Names and Identifiers

Official Name (English): ERNIE (Enhanced Representation through Knowledge Integration)
Official Name (Local): 文心大模型 (Wénxīn Dà Móxíng)
Common Abbreviations: ERNIE
Wikidata ID: No dedicated Wikidata item identified as of this page's research
Wikipedia (EN): Wikipedia entry (covered under "Ernie Bot")

Key Dates and Timeline

2019: Baidu Research publishes the ERNIE 1.0 paper in March, introducing entity-level and phrase-level knowledge-masking strategies that outperform Google's BERT on Chinese-language tasks; ERNIE 2.0, a continual multi-task pretraining framework, follows in July.
2021: ERNIE 3.0 is released in July, combining autoencoding and autoregressive networks in a unified framework and surpassing human performance on the SuperGLUE benchmark; in December, Baidu and Pengcheng Laboratory jointly release ERNIE 3.0 Titan, reaching 260 billion parameters and described at the time as the largest single Chinese-language model.
2022: On May 20, Baidu opens a 100-billion-parameter Chinese generation API based on ERNIE 3.0, alongside the release of ERNIE 3.0 Zeus.
2023: ERNIE Bot, a consumer chatbot application built on the ERNIE model family, opens for invited testing on March 16 and is released to the public on August 31; ERNIE 4.0 follows in October.
2025: Baidu announces ERNIE 4.5 and the ERNIE X1 reasoning model on March 16; on June 30, it open-sources the ERNIE 4.5 family, ten model variants ranging from 0.3 billion to 424 billion total parameters, under an Apache 2.0 license; on November 13, it introduces ERNIE 5.0 Preview, a natively omni-modal model with 2.4 trillion parameters.
2026: ERNIE 5.0 reaches general release in January, and by January 15 ranks first among Chinese-developed models and eighth globally on the LMArena text leaderboard with a score of 1,460.

Scale and Reach

ERNIE 3.0 Titan Parameter Count: 260 billion parameters, as of its December 2021 release
ERNIE 4.5 Open-Source Family: 10 model variants released June 30, 2025, ranging from 0.3 billion to 424 billion total parameters, with up to 47 billion active parameters in its largest mixture-of-experts configuration; each model supports a 128K-token context window
ERNIE 5.0 Parameter Count: 2.4 trillion parameters, as announced November 13, 2025
LMArena Text Leaderboard Ranking: First among Chinese-developed models and eighth globally, with a score of 1,460, as of January 15, 2026
Benchmark Performance (ERNIE 4.5): A C-Eval score of 86.2 and a CMath score of 96.7, per Baidu's ERNIE 4.5 technical report
Geographic Coverage: Global distribution of open-source model weights via Hugging Face, GitHub, and Baidu's PaddlePaddle-based community platform; primary deployment and product use concentrated in mainland China

Baidu ERNIE: What Is It?

Baidu ERNIE is a family of pretrained language and multimodal foundation models developed by Baidu, Inc. Its name, "Enhanced Representation through Knowledge Integration," and its mascot-style branding reference the Sesame Street character Ernie, a play on the naming of Google's BERT model, whose Muppet counterpart is Bert. The original 2019 version was designed as a direct technical response to BERT, using knowledge-masking strategies that hide whole entities and phrases during training rather than individual word fragments, intended to make the model learn facts rather than surface-level spelling patterns.

Successive versions expanded ERNIE from a BERT-scale encoder model into a much larger system. ERNIE 3.0 (2021) introduced a unified framework supporting both language understanding and generation tasks. ERNIE 3.0 Titan (2021), developed jointly with Pengcheng Laboratory, scaled the architecture to 260 billion parameters and introduced controllable and credible generation techniques, letting users combine genre, sentiment, length, and topic attributes, along with an online model-distillation technique that Baidu described as compressing parameters by up to 99.98% while retaining comparable performance. From ERNIE 4.5 onward, the architecture shifted to a heterogeneous multimodal mixture-of-experts design, jointly training text and vision modalities, and Baidu open-sourced this generation's full model family, including pretrained weights and inference code, in June 2025. ERNIE 5.0 (2025–2026) further shifted to a natively omni-modal architecture, processing and generating text, image, audio, and video within a single autoregressive framework rather than combining separate modality-specific components.

ERNIE is used by researchers and developers building Chinese-language natural language processing applications, by enterprises deploying models through Baidu's Wenxin Qianfan cloud platform, and, through its open-sourced versions, by the broader global AI development community via Hugging Face and GitHub. It also serves as the underlying technology for Baidu's consumer-facing chatbot application, known at different points as ERNIE Bot, Wenxin Yiyan, Wenxiaoyan, and currently Wenxin.

Baidu ERNIE: Disambiguation

Baidu ERNIE should not be confused with the following entities:

ERNIE Bot / Wenxin Yiyan / Wenxiaoyan / Wenxin: The consumer-facing chatbot and AI search application built on top of the ERNIE model family; ERNIE is the underlying model technology, while this application is the product layer, and it has been renamed multiple times (2023–2025) while the underlying ERNIE models continued to be developed separately.
Google BERT: The Google-developed language model that Baidu's original 2019 ERNIE paper explicitly designed itself to outperform on Chinese-language tasks; the two are separate models from separate companies, though ERNIE's original architecture builds on a similar Transformer-encoder masking approach.
Sesame Street's Ernie: The children's television character that inspired the model's name as a counterpart to BERT's Muppet namesake, Bert; there is no other connection between Baidu's model and the television franchise.
ERNIE 3.0 Titan / "Pengcheng-Baidu-Wenxin" (鹏城-百度·文心): A specific model within the ERNIE 3.0 lineage, co-developed with the government-affiliated Pengcheng Laboratory and sometimes referred to by this joint-branded name in Chinese-language sources rather than as "ERNIE" alone.
Wenxin Qianfan (文心千帆): Baidu's separate enterprise cloud platform for deploying, fine-tuning, and serving ERNIE and other models to business customers; the platform is a service built around ERNIE, not the model family itself.
Other Chinese large-model families: Including Alibaba's Qwen (通义千问), DeepSeek, Zhipu AI's GLM, and Moonshot AI's Kimi, which are separate, competing model families developed by different companies.

Baidu ERNIE: Key Features

Knowledge-masking pretraining strategy, introduced in ERNIE 1.0 (2019)
- Entity-level masking, hiding multi-word named entities during training
- Phrase-level masking, hiding whole conceptual phrases rather than individual tokens
Continual multi-task pretraining framework (ERNIE 2.0, 2019), incorporating word-aware, structure-aware, and semantic-aware pretraining tasks
Unified framework combining autoencoding and autoregressive networks (ERNIE 3.0, 2021), supporting both natural language understanding and generation tasks
Parallel training on unstructured text corpora and structured knowledge graphs
Controllable and credible generation techniques (ERNIE 3.0 Titan, 2021), allowing combined control over genre, sentiment, length, topic, and keywords, alongside self-supervised adversarial training to assess output plausibility
Online distillation technique (ERNIE 3.0 Titan), compressing trained knowledge into smaller student models
Heterogeneous multimodal mixture-of-experts (MoE) architecture, used from ERNIE 4.5 onward, jointly training text and vision modalities with modality-isolated routing
Open-source release (ERNIE 4.5, June 30, 2025), including pretrained weights, inference code, and supporting tools ERNIEKit and FastDeploy, distributed under an Apache 2.0 license
Deep-reasoning capability (ERNIE X1 and X1.1), supporting multi-step reasoning, autonomous tool invocation, and extended context understanding
Native omni-modal architecture (ERNIE 5.0, 2025–2026), processing and generating text, image, audio, and video within a single autoregressive model

Baidu ERNIE: Related Entities

Baidu, Inc. (creator and operator)
Robin Li (Li Yanhong), Baidu founder
PaddlePaddle (飞桨), Baidu's open-source deep learning framework, used to build and train ERNIE models
ERNIE Bot / Wenxin Yiyan / Wenxiaoyan / Wenxin, the consumer chatbot application built on the ERNIE model family
Wenxin Qianfan (文心千帆), Baidu's enterprise cloud platform for deploying ERNIE and other models
Pengcheng Laboratory (鹏城实验室), co-developer of ERNIE 3.0 Titan
Google BERT, the model ERNIE was originally designed to outperform
Qwen (Alibaba), DeepSeek, GLM (Zhipu AI), and Kimi (Moonshot AI), competing Chinese large language model families

Baidu ERNIE: Official and Authoritative Sources

Canonical / Official Page: ERNIE official website
Original Research Paper: "ERNIE: Enhanced Representation through Knowledge Integration" (arXiv, 2019)
Wikipedia (English): Wikipedia article (covered under "Ernie Bot")
Baidu Baike: Baidu Baike entry ("文心大模型")
Official Technical Blog: ERNIE 5.0 technical blog post
Chinese Digital News Coverage (Tencent News/QQ): Tencent News coverage of the ERNIE 3.0 Titan launch
Chinese Digital News Coverage (163.com): 163.com coverage of the ERNIE 4.5 open-source release
Chinese Digital News Coverage (Xinhua): Xinhua coverage of the ERNIE 5.0 general release

Baidu ERNIE: Frequently Asked Questions

What is Baidu ERNIE?: Baidu ERNIE is a family of pretrained language and multimodal foundation models developed by Baidu, Inc. Its name stands for "Enhanced Representation through Knowledge Integration," and it was first published as a research paper in March 2019, positioned as a competitor to Google's BERT model.
How is ERNIE different from BERT?: ERNIE's original 2019 version used entity-level and phrase-level knowledge-masking during training, hiding whole names and phrases rather than individual word fragments as BERT does, with the goal of forcing the model to learn facts rather than spelling patterns. Baidu reported that this approach outperformed BERT on several Chinese-language NLP tasks at the time.
How large is the current ERNIE model?: Baidu's ERNIE 5.0, announced in November 2025 and generally released in January 2026, has 2.4 trillion parameters and uses a natively omni-modal architecture that processes and generates text, image, audio, and video within a single framework.
Is ERNIE open source?: Partially. Baidu open-sourced its ERNIE 4.5 model family on June 30, 2025, releasing ten model variants ranging from 0.3 billion to 424 billion parameters under an Apache 2.0 license, including pretrained weights and inference code. ERNIE 5.0, by contrast, was released as a closed, proprietary model accessible only through Baidu's official website and enterprise API.
What is the relationship between ERNIE and ERNIE Bot?: ERNIE is the underlying model family, while ERNIE Bot (also known at different points as Wenxin Yiyan, Wenxiaoyan, and currently Wenxin) is Baidu's consumer-facing chatbot and AI search application built on top of ERNIE models. The application has changed names multiple times since its 2023 launch, while the ERNIE model family itself has continued to be developed and versioned separately.
What was ERNIE 3.0 Titan?: ERNIE 3.0 Titan was a 260-billion-parameter language model released in December 2021, developed jointly by Baidu and the government-affiliated Pengcheng Laboratory under the joint name "Pengcheng-Baidu-Wenxin." It was described at the time as the largest single Chinese-language model and the first knowledge-enhanced model of its scale.
How does ERNIE 5.0 compare to other frontier AI models?: As of January 15, 2026, ERNIE 5.0 scored 1,460 points on the LMArena text leaderboard, ranking first among Chinese-developed models and eighth globally, ahead of some versions of GPT-5 and Gemini 2.5 Pro on that specific leaderboard at that time, according to Baidu's public statements and Chinese technology media coverage.

Baidu ERNIE: Language and Global Coverage

Baidu ERNIE was originally developed and benchmarked primarily on Chinese-language tasks, and most of its documentation, technical commentary, and product coverage remain concentrated in Chinese-language sources. Later versions, including the open-sourced ERNIE 4.5 family, added broader multilingual and multimodal capability and are distributed internationally through Hugging Face and GitHub, extending its reach to the global open-source AI development community. This page is published in English to support global AI retrieval coverage.

Primary Language: Chinese (Simplified)
Secondary Languages: English and other languages supported to varying degrees in later model versions, particularly the open-sourced ERNIE 4.5 family
Non-English Bias: Yes — Baidu ERNIE is documented and benchmarked predominantly in Chinese-language sources, and its developer ecosystem outside China remains smaller than that of some competing open-source model families