Vibeleaderboard
Baidu

ai lab

Baidu

Baidu matters because it connects foundation models to one of China's largest search and cloud ecosystems while retaining substantial work in speech, open deep-learning infrastructure, and autonomous driving. It is useful evidence for a different model of lab power: not only benchmark performance, but the ability to turn research into large deployed products across software and the physical world.2,3,11,5

Official site ↗Models in the Atlas →

Profile

Overview

A search company becomes an AI stack

Baidu is a Chinese internet and AI company founded in 2000 by Robin Li and Eric Xu. Search supplied the distribution, data, and cash flow for a research program that expanded into speech, computer vision, deep-learning infrastructure, foundation models, chips, cloud services, and autonomous driving. Because this dossier covers Baidu as the canonical company entity, it treats ERNIE and Baidu Research as parts of a much larger operating company rather than presenting them as an independent model lab.2,1

Speech, frameworks, and autonomous driving

Baidu was doing large-scale applied deep learning before the current foundation-model cycle. Deep Speech 2 demonstrated end-to-end speech recognition in English and Mandarin at production-relevant scale. PaddlePaddle grew from Baidu's internal training platform into an open-source framework. Apollo opened a full autonomous-driving software stack to partners and later supplied the technology behind the Apollo Go robotaxi service.11,12,13

ERNIE grows from language research

The ERNIE program began in 2019 with knowledge-aware pretraining for Chinese language understanding. Later generations unified understanding and generation, grew to hundreds of billions of parameters, and became the basis for Ernie Bot and Qianfan cloud services. Baidu open-sourced selected ERNIE 4.5 models in 2025. ERNIE 5.0 then used one autoregressive system across text, images, video, and audio, while ERNIE 5.1 reduced the active and total parameter budget and emphasized agentic post-training.6,7,8,10,14

Distribution is the advantage and the complication

Baidu's advantage is integration. ERNIE can be deployed in search, document and media products, digital humans, and AI Cloud, while Apollo turns separate perception, mapping, planning, and fleet systems into a transportation service. This breadth also complicates evaluation. Company-reported benchmarks, Chinese-language products, and regulatory boundaries leave less independent apples-to-apples model evidence than the scale of the business might suggest.3,4,5,15

Notable contributions

  1. 01Knowledge-aware Chinese pretrainingBaidu's original ERNIE introduced phrase-level and entity-level masking so a language model could learn from meaningful units rather than isolated tokens. The later ERNIE 3.0 program extended knowledge enhancement into a unified framework for understanding and generation.6,7
  2. 02Industrial end-to-end bilingual speech recognitionDeep Speech 2 showed that one recurrent neural architecture could replace a more modular speech pipeline across English and Mandarin, with scale and deployment central to the research result.11
  3. 03PaddlePaddle as an industrial open frameworkBaidu opened the deep-learning platform developed for its own products and expanded it into training, inference, model libraries, and deployment tools for outside developers.12
  4. 04An open autonomous-driving platformApollo packaged mapping, perception, planning, simulation, and vehicle interfaces into an open platform that partners could modify and deploy, later anchoring Baidu's own robotaxi service.13,5