AI Atlas AI ATLAS

ERNIE

A Chinese model that began with knowledge-enhanced pretraining, an early landmark version

Baidu Model ClosedLegacy

inTextText

WHAT IT IS

ERNIE is Baidu’s large language model series; the earliest, ERNIE 1.0, was released in March 2019 and stands for Enhanced Representation through Knowledge Integration. Its key idea is knowledge-enhanced pretraining: entity- and phrase-level masking during pretraining forces the model to learn knowledge relations between words, not just surface co-occurrence of neighbouring characters. The ERNIE line later grew into a large family spanning retrieval, dialogue and multimodality, and underpins Baidu’s ERNIE Bot product. This entry records the ERNIE starting point as represented by the early 2019 versions.

Why it matters

It introduced entity- and phrase-level masking into pretraining so the model learns knowledge relations rather than mere character co-occurrence — an idea widely borrowed in early Chinese pretraining. These early versions have since been superseded by later ERNIE generations.

Key specs

Architecture: Knowledge-enhanced pretraining (knowledge masking)
Released: 2019-03 (ERNIE 1.0)
Modality: Text
Open weights: No

Capabilities

Text Generation

Continue a passage, one word at a time

Conversation & Instruction Following

Understand intent across turns and act on it

Question Answering & RAG

Retrieve the evidence first, then answer from it

Related concepts

Pretraining & Fine-tuning

Learn language first from vast unlabelled text, then specialise with little data — the most data-efficient paradigm in modern AI

NLP & Large Language Models

Transformer Architecture

Replacing word-by-word relay with a room where everyone speaks at once, so long-range dependencies are one hop away

NLP & Large Language Models

Tokenization

Models do not read characters, they read tokens — and how you split text quietly sets both capability and cost

NLP & Large Language Models

Comparable products

GLM

A Chinese general model that began with autoregressive blank-infilling pretraining

Qwen

An open-weight family spanning many sizes, with multimodal versions

Model Open weights

Hunyuan

Tencent (Hunyuan)

Tencent’s general model family, with open-weight versions

Model Open weights

Baichuan

An open-weight general model aimed at Chinese-language use

Model Open weights

Doubao

ByteDance (Seed)

ByteDance’s general chat model and application

Pangu

Huawei’s Pangu family of foundation models