Baidu AI

RawGraph

Last edited

Fact-checked

In review queue

Sources

80 citations

Revision

v5 · 6,778 words

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Baidu AI
百度人工智能
TypeDivision/Business Unit
IndustryArtificial intelligence, Machine learning, Deep learning
Founded2013 (Early AI research)**2017 (AI Group formation)
HeadquartersBeijing, China
Key peopleRobin Li (CEO)
Wang Haifeng (CTO, Head of AIG)
ParentBaidu, Inc.
ProductsERNIE (Foundation models)
Apollo (Autonomous driving)
DuerOS (Conversational AI)
PaddlePaddle (Deep learning framework)
Xiaodu (Smart devices)
Kunlunxin AI chips
RevenueRMB 7.1 billion (AI Cloud Q4 2024)
Employees2,000+ (AIG division)
Websiteresearch.baidu.com

Baidu AI is the artificial intelligence division of Baidu, Inc. that pursues a full-stack AI strategy spanning chips, deep learning frameworks, foundation models, and applications. Its four pillars are the ERNIE (Enhanced Representation through Knowledge Integration) family of large language models and the Ernie Bot assistant, Apollo and Apollo Go autonomous driving, the Kunlunxin AI accelerator chips, and the open-source PaddlePaddle framework. In fiscal year 2025, CEO Robin Li declared that "2025 marked a pivotal year as AI became the new core of Baidu," with revenue from Baidu Core's AI-powered business reaching RMB 40.0 billion, up 48% year over year.[1][78][79]

Baidu AI operates through the Artificial Intelligence Group (AIG) and provides AI technologies that power hundreds of millions of daily interactions across Baidu's ecosystem.[1]

What is Baidu's AI strategy?

Baidu's strategy is to own every layer of the AI stack and reuse it across its consumer and enterprise products. The same chips (Kunlunxin), framework (PaddlePaddle), models (ERNIE), and cloud (Qianfan) that train Baidu's own systems are sold to external developers and enterprises, while ERNIE rebuilds Baidu's existing products from search to maps. Robin Li framed the goal at Baidu World 2025: "When you internalize AI, it becomes a native capability and transforms intelligence from a cost into a source of productivity."[78] This vertical integration distinguishes Baidu from Chinese rivals such as Alibaba, ByteDance, and Tencent, none of which combine frontier models, in-house AI silicon, a national-scale cloud, and the largest robotaxi fleet in a single company.[68][79]

Overview

Baidu AI spans fundamental research, platforms, and applied products across multiple domains:

  • Foundation models & Generative AI: The ERNIE (Enhanced Representation through Knowledge Integration) series are transformer-based foundation models. The public-facing chatbot Ernie Bot (Wenxin Yiyan) reached over 100 million users by December 2023, 200 million users by April 2024, and 300 million users by June 2024. ERNIE Bot was made free to use from April 1, 2025.[2][3]

  • AI Cloud: Baidu AI Cloud provides model-as-a-service, training/inference, and industry solutions through its Qianfan platform. The service reported 26% year-over-year revenue growth in Q4 2024 and 42% growth in Q1 2025. AI Cloud infrastructure revenue reached approximately RMB 20 billion for the full year 2025, up 34% year over year.[4][5][70]

  • Autonomous driving: Under Apollo, Baidu operates Apollo Go robotaxis across 22 cities worldwide with a 1,000-vehicle global fleet. The service completed over 17 million cumulative rides by November 2025, with weekly rides reaching 250,000, matching Waymo's volume.[6][7]

  • Developer ecosystem: PaddlePaddle, open-sourced in 2016, serves as China's leading deep-learning framework. As of April 2025, the ERNIE and PaddlePaddle developer ecosystem had grown to 21.85 million developers serving 670,000 enterprises, with over 1.1 million models created.[8][9]

  • AI hardware: Through Kunlunxin, Baidu designs AI accelerators. The third-generation P800 Kunlun chip entered mass production, and Baidu deployed a 30,000-chip cluster capable of training models with hundreds of billions of parameters. Kunlunxin filed confidentially for a Hong Kong IPO in January 2026.[10][11]

Robin Li and the Founding of Baidu

Robin Li (Li Yanhong, born November 17, 1968, in Yangquan, Shanxi Province, China) is the co-founder, chairman, and CEO of Baidu, Inc. He studied information management at Peking University before earning a master's degree in computer science from the University at Buffalo in the United States.[12]

Li's career in search technology began in 1994 when he joined IDD Information Services, a New Jersey division of Dow Jones and Company. There he helped develop software for the online edition of The Wall Street Journal. In 1996, while at IDD, Li invented the RankDex site-scoring algorithm, which used hyperlinks to measure website quality. This was the first search engine to rank pages based on link analysis. Google co-founder Larry Page cited Li's RankDex work in his 1998 patent filing for the PageRank algorithm.[12][13]

After working at Infoseek from 1997 to 1999, where he developed an image search function used by Go.com, Li returned to China and co-founded Baidu with Eric Xu in January 2000. The company rapidly grew into China's dominant search engine, commanding roughly 80% market share in Chinese-language search. Baidu listed on NASDAQ on August 5, 2005.[12]

Li has been a consistent advocate for AI investment within Baidu. He has described the company's strategy as rebuilding "almost every one of their existing products" on top of ERNIE foundation models, and has pledged to increase AI spending to ensure Baidu remains at the forefront of technological innovation. In early 2025, Li predicted an "exponential boom" in AI applications, noting that "the seeds we planted in 2023 and 2024 will hopefully take root, blossom and bear fruit."[14]

History and Development

Early Development (2013-2016)

Baidu began its serious investment in artificial intelligence in 2013, establishing the Institute of Deep Learning (IDL) as its first dedicated AI research laboratory.[15] The company was among the first Chinese technology companies to recognize the potential of deep learning, developing capabilities shortly after Geoffrey Hinton's breakthrough at the ImageNet competition in 2012.[8]

In 2014, Baidu hired Andrew Ng, co-founder of Google Brain and a prominent AI researcher from Stanford University, as its Chief Scientist to lead Baidu Research and the company's AI initiatives.[16] Under Ng's leadership, Baidu Research was established with labs in Silicon Valley and Beijing, and the AI group grew to approximately 1,300 people by 2017, including 300 researchers in Baidu Research.[16]

Formation of AI Group (2017-2018)

In January 2017, Baidu recruited former Microsoft executive Qi Lu as Chief Operating Officer to oversee its AI strategy. Under Lu's leadership, Baidu formally consolidated its various AI units into the Artificial Intelligence Group (AIG) in March 2017.[17]

Following Andrew Ng's departure in March 2017, Wang Haifeng was appointed to lead the expanded AI division, overseeing machine learning, big data, computer vision, natural language processing, speech technology, knowledge graph, and augmented reality.[18][19]

In May 2018, following Qi Lu's departure, Wang Haifeng was promoted to Senior Vice President and General Manager of AIG.[20] In May 2019, he was further promoted to Chief Technology Officer.[21]

Recent Developments (2019-2025)

In 2018, China's government designated Baidu as one of its "AI champions," recognizing the company's leadership in artificial intelligence development.[1] The company has continued to expand its AI capabilities, with the AIG growing to over 2,000 scientists and engineers.[22]

In March 2023, Baidu launched Ernie Bot (Wenxin Yiyan), one of China's first generative AI chatbots, competing with global models like ChatGPT.[23] The bot opened to the general public on August 31, 2023, after regulatory approval. By December 2023, Ernie Bot had surpassed 100 million users, making it one of the fastest-growing AI applications in the world.[23][24]

In 2024, the pace of development accelerated. Baidu released ERNIE 4.0 Turbo in June, and Ernie Bot user counts reached 200 million by April and 300 million by June. The company also made ERNIE Speed and ERNIE Lite free for all users as part of an industry-wide price war in China's AI market.[25][26]

In March 2025, Baidu released two new AI models: ERNIE 4.5, a multimodal foundation model, and ERNIE X1, a reasoning model, claiming performance on par with DeepSeek R1 at half the price. In June 2025, the ERNIE 4.5 model family was open-sourced under the Apache 2.0 license. At Baidu World 2025 on November 13, the company unveiled ERNIE 5.0, a 2.4-trillion-parameter natively omni-modal model.[3][27][28]

Timeline of Major Milestones

DateMilestoneDescription
2000Baidu foundedRobin Li and Eric Xu co-found Baidu in Beijing[12]
2013Institute of Deep Learning establishedBaidu's first dedicated AI research lab focusing on deep learning technologies[15]
2014Andrew Ng joins as Chief ScientistAccelerated AI research and established Baidu Research[16]
2016PaddlePaddle open-sourcedReleased open-source deep learning framework to developers[8]
2016Baidu Brain unveiledLaunched core AI engine and technology platform[29]
January 2017Qi Lu joins as COOFormer Microsoft executive recruited to oversee AI strategy[17]
March 2017AI Group (AIG) formedConsolidated AI units under unified structure[18]
2017Apollo platform launchedOpen-source autonomous driving platform introduced[30]
2018China AI champion designationGovernment recognized Baidu as national AI leader[1]
March 2019ERNIE 1.0 releasedFirst version of knowledge-enhanced language model[31]
December 2021ERNIE 3.0 launchedScaled to 10 billion parameters with unified understanding and generation[32]
August 2022Apollo Go fully driverlessFirst permits for fully driverless robotaxi services in China[33]
March 2023Ernie Bot public launchGenerative AI chatbot based on ERNIE models[23]
August 31, 2023Ernie Bot opens to publicGeneral availability after regulatory approval[23]
December 2023100 million Ernie Bot usersMajor adoption milestone in four months[24]
October 2023ERNIE 4.0 releasedClaimed performance comparable to GPT-4[34]
April 2024200 million Ernie Bot usersUser base doubled in under four months[2]
May 2024ERNIE Speed and Lite made freeTwo lightweight models offered at no cost for developers[25]
June 2024ERNIE 4.0 Turbo releasedFaster responses, improved reasoning, 300 million users reached[26]
March 2025ERNIE 4.5 & X1 releaseMultimodal and reasoning models, Ernie Bot made free[3]
June 2025ERNIE 4.5 open-sourcedTen model variants released under Apache 2.0 license[27]
Q2 20252.2 million Apollo Go driverless ridesQuarterly record for fully autonomous rides[7]
August 2025Kunlun chip mass productionThird-generation P800 AI chips in production[10]
September 2025Apollo Go Dubai permitsFirst autonomous driving deployment in the Middle East[35]
November 13, 2025ERNIE 5.0 unveiled2.4-trillion-parameter omni-modal model at Baidu World 2025[28]
January 2026Kunlunxin Hong Kong IPO filingAI chip subsidiary files confidentially for listing[11]

Core Technologies

ERNIE Foundation Models

ERNIE (Enhanced Representation through Knowledge Integration) is Baidu's series of large language models that has been in development since 2019.[31] The ERNIE family represents Baidu's core generative AI technology, with models evolving from knowledge-enhanced pre-training to multimodal capabilities. The name reflects the model's foundational approach of integrating structured knowledge (from knowledge graphs and entity-level masking) into the pre-training process, distinguishing it from models that rely solely on raw text.

Comprehensive ERNIE Model Family

ModelRelease DateParametersArchitectureKey Features
ERNIE 1.0March 2019Not disclosedTransformerIntroduced knowledge integration by masking entities and phrases, not just single words[31]
ERNIE 2.0July 2019Not disclosedTransformerContinual pre-training framework with multi-task learning across lexical, syntactic, and semantic tasks[36]
ERNIE 3.0December 202110 billionTransformer-XLUnified framework excelling at both language understanding and generation; 48-layer universal representation module with 4096 hidden units[32]
ERNIE 3.52023Not disclosedTransformer-XL2x training throughput and 17x inference throughput vs ERNIE 3.0; knowledge snippet enhancement; surpassed ChatGPT (3.5) in composite ability scores[37]
ERNIE 4.0October 2023Not disclosedTransformerFull upgrade in understanding, generation, reasoning, and memory; claimed performance comparable to GPT-4[34]
ERNIE 4.0 TurboJune 2024Not disclosedTransformerFaster responses and boosted reasoning; 70% cheaper than ERNIE 4.0; input: 0.03 yuan/1K tokens[26]
ERNIE SpeedMarch 2024Not disclosedDenseHigh-performance lightweight model for fine-tuning; 8K and 128K context; made free May 2024[25]
ERNIE LiteMarch 2024Not disclosedDenseLightweight model balancing performance and inference speed; optimized for low-power hardware; made free May 2024[25]
ERNIE 4.5March 2025Up to 424B total (47B active, MoE)Mixture-of-ExpertsNative multimodal (text, image, audio, video); 10 model variants; open-sourced under Apache 2.0 in June 2025[3][27]
ERNIE X1March 2025Not disclosedTransformerFirst reasoning model with multimodal capabilities; performance on par with DeepSeek R1 at half the price[3]
ERNIE 4.5 TurboApril 2025Not disclosedMoECost-optimized version of ERNIE 4.5; 80% cheaper than competitors[38]
ERNIE X1 TurboApril 2025Not disclosedTransformerFaster reasoning variant with reduced latency[38]
ERNIE 5.0November 20252.4 trillion total (MoE, under 3% experts activated per inference)Omni-modal MoENatively omni-modal; jointly models text, images, audio, and video; excels in agentic planning and tool use[28][80]

ERNIE 3.0 and 3.5

ERNIE 3.0, released in December 2021, marked a significant step in scaling Baidu's language models. Trained on a 4TB corpus of plain text and a large-scale knowledge graph, the model reached 10 billion parameters. Its architecture features a 48-layer universal representation module with 4,096 hidden units and 64 attention heads, paired with task-specific modules of 12 layers each. ERNIE 3.0 used the Transformer-XL backbone and fused auto-regressive and auto-encoding approaches, enabling the same pre-trained model to handle both understanding and generation tasks through zero-shot, few-shot, or fine-tuned learning.[32]

ERNIE 3.5 improved on this foundation with a 2x increase in training throughput and a 17x increase in inference throughput compared to ERNIE 3.0. It introduced "Knowledge Snippet Enhancement," where the model analyzes user queries, identifies relevant knowledge snippets from knowledge graphs and search engines, and uses those snippets to generate more accurate responses. ERNIE 3.5 surpassed ChatGPT (3.5) in composite ability scores and outperformed GPT-4 in several Chinese-language benchmarks.[37]

ERNIE 4.0 and 4.0 Turbo

ERNIE 4.0 was released in October 2023 alongside the broader public launch of Ernie Bot. Baidu described it as a full upgrade across four dimensions: understanding, generation, reasoning, and memory. The company claimed its performance was comparable to GPT-4 across multiple benchmarks.[34]

ERNIE 4.0 Turbo followed in June 2024, offering faster response times and enhanced reasoning at 70% lower cost than the original ERNIE 4.0. Input pricing was set at 0.03 yuan per thousand tokens, with output at 0.06 yuan per thousand tokens. The Turbo release coincided with Ernie Bot reaching 300 million users.[26]

ERNIE 4.5 and Open-Source Release

Released in March 2025, ERNIE 4.5 is Baidu's first native multimodal model, capable of understanding and generating text, images, audio, and video content.[3] The model family uses a Mixture-of-Experts architecture with a novel heterogeneous modality structure that supports parameter sharing across modalities while also allowing dedicated parameters for each modality. The largest variant has 424 billion total parameters with 47 billion active parameters.

On June 30, 2025, Baidu open-sourced the full ERNIE 4.5 family (10 model variants) under the Apache 2.0 license, making them available on Hugging Face, GitHub, and the PaddlePaddle ecosystem. The variants span large-scale MoE models with 47 billion and 3 billion active parameters (up to 424 billion total) down to dense models as small as 0.3 billion parameters. The models were trained using PaddlePaddle with a model FLOPs utilization (MFU) of up to 47% for the largest language model. Input pricing for the API was set at RMB 0.004 per thousand tokens, roughly 1% of GPT-4.5's cost.[27]

ERNIE X1

ERNIE X1 is Baidu's first reasoning model with multimodal capabilities, designed for deep-thinking tasks.[3] Features include:

  • Performance comparable to DeepSeek R1 at half the price
  • Support for advanced search, document Q&A, image understanding, and code interpretation
  • Technologies including Progressive Reinforcement Learning and End-to-End Training

ERNIE 5.0

Unveiled at Baidu World 2025 on November 13, 2025, ERNIE 5.0 is a 2.4-trillion-parameter natively omni-modal foundation model that jointly models text, images, audio, and video.[28][80] It uses a Mixture-of-Experts design that activates less than 3% of its experts on each inference pass to hold down GPU demand and latency, and it is built on the PaddlePaddle framework.[80] The model delivers upgraded capabilities in multimodal understanding, instruction following, creative writing, factual reasoning, agentic planning, and tool use. ERNIE 5.0 is available through Ernie Bot for consumers and through Baidu AI Cloud's Qianfan platform for enterprise users.[28]

On the public LMArena leaderboard published January 15, 2026, the Ernie-5.0-0110 checkpoint scored 1,460 points, ranking 8th globally and 1st among all Chinese models, putting it on par with OpenAI's GPT-5.1 (High) and ahead of Google's Gemini 2.5 Pro and Anthropic's Claude Sonnet 4.5.[80]

Ernie Bot (Wenxin Yiyan)

Ernie Bot (Chinese: 文心一言; pinyin: Wenxin Yiyan) is Baidu's consumer-facing generative AI chatbot, launched in March 2023 as one of China's first responses to ChatGPT. The chatbot is powered by the ERNIE family of models and is available via web, mobile app, and API.

User growth milestones:

DateUsersNotes
March 20231.2 millionInitial waitlist signups[23]
August 31, 2023General availabilityPublic launch after regulatory approval[23]
December 2023100 millionReached in approximately four months after public launch[24]
April 2024200 millionDoubled in under four months[2]
June 2024300 millionCoincided with ERNIE 4.0 Turbo release[26]
April 2025Free tier launchedErnie Bot made free for consumers starting April 1, 2025[3]

By December 2024, Ernie Bot's API was processing 1.65 billion calls daily, a 33x increase over the course of the year. The chatbot supports 14.65 million developers building applications on the platform.[39][40]

PaddlePaddle Deep Learning Framework

PaddlePaddle (PArallel Distributed Deep LEarning, Chinese: 飞桨; pinyin: Feijiang) is Baidu's open-source deep learning platform, first developed in 2013 and open-sourced in 2016.[8] It serves as a direct competitor to global frameworks like Google's TensorFlow and Meta's PyTorch.

Key statistics and features:

  • As of April 2025, the combined ERNIE and PaddlePaddle ecosystem has attracted 21.85 million developers[9]
  • Serving 670,000 enterprises
  • Over 1.1 million models generated
  • Ranks first in China in deep learning framework application scale
  • PaddlePaddle 3.0, released in 2025, was redesigned specifically for large model development and inference[41]

Comprehensive toolsets include:

  • PaddleDetection (object detection)
  • PaddleSeg (semantic segmentation)
  • PaddleHub (pre-trained model hub)
  • PaddleNLP (natural language processing)
  • PaddleOCR (optical character recognition)
  • PaddleSpeech (speech recognition and synthesis)[42]

Baidu Brain

Baidu Brain is the company's core AI engine and technology platform, first unveiled in 2016.[29] It serves as the foundational infrastructure for all of Baidu's AI-related businesses and integrates:

  • Computer vision: Image recognition, facial recognition, and OCR
  • Natural Language Processing: Language understanding, sentiment analysis, machine translation
  • Voice Technology: Speech recognition and speech synthesis
  • Knowledge Graph: Large-scale structured knowledge base
  • Deep Learning: Underlying frameworks and models

As of 2022, Baidu Brain evolved to version 7.0, with over 1,397 technical capabilities used by 8 million developers.[43]

Products and Services

Apollo Autonomous Driving

Apollo is Baidu's autonomous driving platform, launched in 2013 with the open platform introduced in 2017.[44][30]

Apollo Open Platform

An open-source autonomous driving platform providing:

  • Complete hardware and software service solutions
  • Cloud data services
  • Testing tools and development frameworks
  • The Apollo ADFM (Autonomous Driving Foundation Model) large model for driving decisions
  • Version 9.0 with enhanced development and debugging experience[30]

What is Apollo Go and how does it compare to Waymo?

Apollo Go is Baidu's robotaxi service and one of the world's largest autonomous ride-hailing operations, competing directly with Waymo in scale. As of October 31, 2025, Apollo Go was completing over 250,000 fully driverless rides per week, on par with the weekly volume Waymo disclosed earlier in 2025, and had logged more than 17 million cumulative ride orders, 240 million total autonomous kilometers, and 140 million fully driverless kilometers.[28][46]

MetricValueDate
Cumulative rides17+ millionNovember 2025[45]
Weekly ride volume250,000+ (matching Waymo)October 2025[46]
Total autonomous distance240+ million km (140+ million fully driverless)November 2025[28]
Q4 2025 driverless rides3.4 million (up over 200% YoY)Q4 2025[79]
Fleet size1,000+ vehicles globallyMay 2025[6]
Cities of operation22 (expanding to 26 by early 2026)2025-2026[45]
Fully driverless transitionFebruary 2025Safety drivers removed from all vehicles[45]
Safety record10.14 million km per airbag deploymentSurpasses human driver averages[47]

Apollo Go transitioned to fully driverless operations across China in February 2025, removing safety drivers from its vehicles. The service completed 2.2 million fully driverless rides in Q2 2025 alone, a 148% year-over-year increase, and 3.4 million fully driverless rides in Q4 2025, up more than 200% year over year.[7][79]

Vehicle generations:

GenerationCostKey Features
5th generation~$48,000Earlier platform with safety driver
6th generation (RT6)~$27,670 (200,000 yuan)Purpose-built robotaxi; 60% cost reduction; launched at Apollo Day 2024[48]
7th generationUnder $20,000 (planned)Further cost reduction targeting profitability[49]

International expansion:

Apollo Go signed a strategic cooperation agreement with Dubai's Roads and Transport Authority (RTA) in March 2025, deploying 100 fully autonomous RT6 vehicles in urban Dubai by the end of 2025, with plans to scale to 1,000 vehicles by 2028. Baidu also announced a partnership with Uber to integrate Apollo Go into Uber's platform in Dubai. In Hong Kong, Apollo Go expanded open-road testing to Kowloon and Kwun Tong District, treating it as a strategic entry point for right-hand-drive markets across Asia, Africa, and the UK. The company uses an asset-light strategy for international expansion, partnering with local operators rather than directly owning fleets.[35][50]

AI Cloud Services and Qianfan Platform

Baidu AI Cloud (百度智能云) provides enterprise AI services including:

  • Infrastructure as a Service (IaaS)
  • Platform as a Service (PaaS)
  • Software as a Service (SaaS)
  • Industry-specific solutions for finance, manufacturing, transportation

Qianfan Model-as-a-Service Platform:

Qianfan is Baidu's flagship model-as-a-service (MaaS) platform, serving as the commercial interface for enterprise customers to access, fine-tune, and deploy ERNIE models. The platform has been upgraded to an agent-centric architecture supporting multimodal and reasoning AI applications. Key Qianfan metrics:

  • Over 85,000 enterprise clients
  • 190,000 AI applications built on the platform
  • Expanded model library including third-party models alongside ERNIE
  • Support for complex agent workflows and tool integration
  • Full-stack offerings from model training to deployment[51]

In Q3 2025, AI Cloud infrastructure revenue reached RMB 4.2 billion (up 33% year over year), while subscription-based AI accelerator infrastructure revenue surged 128%, indicating a shift toward recurring, inference-heavy workloads. That momentum continued into Q4 2025, when AI Cloud infrastructure revenue reached RMB 5.8 billion and subscription-based AI accelerator infrastructure revenue rose 143% year over year.[52][79]

DuerOS Conversational AI Platform

DuerOS is Baidu's conversational AI system launched at Baidu World Conference in September 2015.[53]

Key metrics and capabilities:

  • Conversational skills in 10 major domains and over 200 sub-domains[54]
  • Install base: 400+ million devices (June 2019)[55]
  • Monthly voice queries: 3.6+ billion (excluding Baidu apps)[55]
  • Ecosystem: 200+ partners, 110+ smart device integrations[56]

Smart Devices Division

Xiaodu

Xiaodu is Baidu's smart device brand, launched in 2018.[57]

Products include:

  • Smart Speakers and Displays: Xiaodu Smart Speaker series and Xiaodu Smart Display X10 (flagship with 10.1-inch screen, RMB 999)[58]
  • Sales growth: 384% (2018-2019); ranked third globally after Amazon and Google (2019)[57]
  • XiaoduPods: True wireless earbuds with AI capabilities, 28-hour battery life[59]
  • Xiaodu Qinghe: AI tutor smartphone for children (2023)[60]

AI Hardware

What are Baidu's Kunlun chips?

Kunlunxin (昆仑芯) is Baidu's AI chip subsidiary developing AI accelerators for training and inference workloads. The unit has grown from an internal project into a standalone business preparing for public listing, with Baidu holding a 57.67% controlling stake.[11]

Chip generations and roadmap:

ChipGenerationProcess NodePerformanceStatus
Kunlun 11st gen14nm256 TOPS (INT8)Mass production since 2020
Kunlun 22nd gen7nmImproved over 1st genMass production since 2023
Kunlun P8003rd genAdvanced node~345 TFLOPS (FP16); comparable to Nvidia A100 and Huawei 910BMass production 2025[10]
Kunlun M100Next genTBDOptimized for large-scale inferencePlanned early 2026[61]
Kunlun M300Next genTBDTraining and inference for ultra-large multimodal modelsPlanned 2027[61]
Kunlun N-seriesFutureTBDNext-generation architecturePlanned 2029[61]

In April 2025, Baidu unveiled a 30,000-chip P800 cluster capable of training "DeepSeek-like" models with hundreds of billions of parameters. The cluster serves over 100 enterprise customers. In August 2025, Kunlunxin secured orders worth over RMB 1 billion (US$139 million) from China Mobile.[10][62]

Baidu also announced two "supernode" solutions linking multiple chips: the Tianchi 256 and Tianchi 512, composed of P800 chips, with availability in H1 and H2 2026 respectively. The company's long-term roadmap targets a one-million-card Kunlun single-cluster deployment by 2030.[61]

IPO plans: On January 1, 2026, Kunlunxin filed confidentially for a Hong Kong stock exchange listing to tap surging AI investment. The unit was previously valued at about 21 billion yuan (~$3 billion) during its last fundraising round; Baidu, which holds 57.67% of the company, said a carve-out listing would better reflect Kunlunxin's value to investors focused on general-purpose AI computing chips. Following the proposed spin-off, Kunlunxin is expected to remain a Baidu subsidiary.[11]

Research Organization

Baidu Research

Baidu Research is the company's AI research division with locations in Beijing, Silicon Valley, and Seattle.[63]

Research Laboratories

LaboratoryEstablishedFocus Area
Institute of Deep Learning (IDL)2013Deep learning research[15]
Big Data Lab (BDL)2014Big data technologies and applications
Silicon Valley AI Lab (SVAIL)2014US-based AI research[64]
Business Intelligence Lab (BIL)2018Data analysis technology[65]
Robotics and Autonomous Driving Lab (RAL)2018Computer vision and autonomous driving[65]

National Engineering Laboratory

In 2017, China's National Development and Reform Commission approved Baidu to lead the National Engineering Laboratory of Deep Learning Technology and Application.[66]

Focus areas:

  • Deep learning
  • Computer vision and auditory technology
  • Biometric identification
  • Human-computer interaction
  • Standardization service
  • Deep learning intellectual property

Competitive Landscape

Baidu AI operates in an intensely competitive environment within China's AI industry, facing off against other Chinese tech giants as well as global players.

China AI Cloud Market Share

The Chinese market for AI cloud services expanded rapidly, growing from 20.83 billion yuan in 2024 to an estimated 51.8 billion yuan (US$7.3 billion) in 2025. Market share varies by measurement methodology:

CompanyAI Cloud Market Share (2025 est.)Key AI Products
Alibaba Cloud~35.8%Qwen model series, Tongyi Qianwen chatbot
ByteDance (Volcano Engine)~14.8%Doubao model, Volcano Engine cloud
Huawei Cloud~13.1%Pangu models, Ascend chips
Tencent Cloud~7%Hunyuan model series
Baidu AI Cloud~6.1%ERNIE models, Qianfan platform

Note: In China's public cloud market specifically for AI, Baidu and Alibaba each held roughly a quarter of the market according to some analysts.[67][68]

Price War

Starting in 2024, Chinese tech firms engaged in aggressive LLM pricing competition. Alibaba cut prices by as much as 97% on generative AI models, prompting Baidu to make ERNIE Speed and ERNIE Lite free in May 2024. By March 2025, Baidu positioned ERNIE 4.5 at approximately 1% of the cost of comparable Western models, and in April 2025, Ernie Bot itself was made free for consumers.[25][3]

Model Adoption

In terms of LLM usage among Chinese models, Alibaba's Qwen series led with roughly 17.7% usage share, followed by ByteDance's Doubao at 14.1% and DeepSeek at 10.3%. Baidu differentiates itself through its full-stack approach (combining models, chips, cloud, and applications) and its leadership in autonomous driving, an area where none of its direct Chinese AI cloud competitors operate at comparable scale.[68]

Leadership

  • Robin Li - Co-founder, Chairman, and CEO of Baidu. Born November 17, 1968. Inventor of the RankDex algorithm. Peking University (B.S.) and University at Buffalo (M.S.).[12]

  • Wang Haifeng - Chief Technology Officer, Head of AIG and Baidu Research

    • Joined Baidu in 2010, promoted to Vice President in 2013
    • Senior Vice President since May 2018
    • CTO since May 2019[21]
    • ACL Fellow (2016), IEEE Fellow (2022)[69]
    • Published over 100 academic papers and holds over 100 patents[69]

Financial Performance

Baidu's full-year 2025 results, reported in February 2026, show a company being reshaped by AI: total revenue declined 3% to RMB 129.079 billion as the legacy online advertising business softened, but revenue from Baidu Core's AI-powered business grew 48% to RMB 40.0 billion and reached 43% of Baidu General Business revenue in Q4 2025, up from 26% in fiscal year 2024.[78][79]

MetricPeriodValueYoY Growth
Total RevenueFY 2025RMB 129.079 billion-3%[79]
AI-Powered Business RevenueFY 2025RMB 40.0 billion+48%[79]
AI Cloud Infra RevenueFY 2025~RMB 19.8 billion+34%[79]
AI Cloud + AI Apps RevenueFY 2025RMB 30 billionN/A[79]
AI Applications RevenueFY 2025RMB 10.2 billion+5%[79]
AI Cloud RevenueQ4 2024RMB 7.1 billion+26%[4]
AI Cloud RevenueQ1 2025N/A+42%[5]
AI Cloud Infra RevenueQ3 2025RMB 4.2 billion+33%[52]
AI Cloud Infra RevenueQ4 2025RMB 5.8 billionN/A[79]
Baidu Core RevenueQ1 2025RMB 25.5 billion+7%[5]
Baidu Core RevenueFY 2024RMB 104.7 billion+1%[71]
Total RevenueFY 2024RMB 133.1 billion ($18.24 billion)N/A[71]
Core Business ProfitsFY 2024RMB 23.4 billion+21%[71]
AI-Powered Business RevenueQ4 2025RMB 11+ billion (43% of Baidu General Business)N/A[79]
AI-native Marketing Services RevenueFY 2025N/A+301%[79]
ERNIE Bot API CallsDecember 20241.65 billion daily33x increase YoY[39]
GenAI RevenueFY 2024N/A~3x increase YoY[71]

Baidu's financial trajectory shows a company in active transformation. While total revenue declined modestly in 2025 due to weakness in the online advertising business, AI-related revenue grew rapidly. AI Cloud infrastructure revenue reached roughly RMB 19.8 billion in 2025, up 34% year over year, and AI applications revenue exceeded RMB 10 billion. CFO Haijian He noted that "revenue from Baidu Core AI-powered Business exceeded RMB 11 billion in Q4," and AI-native marketing services revenue grew 301% year over year for the full year.[79]

Applications and Industry Solutions

Baidu AI provides specialized solutions across various sectors:

CategoryProductsDescription
Identity VerificationFace Real-Name AuthenticationRemote identity verification using face recognition for finance and government[72]
Customer ServiceIntelligent Customer ServiceAI-powered service across voice and digital channels[72]
Digital HumansIntelligent Digital Human (Xiling)Service and performance digital humans for finance, media, entertainment[72]
ConferencingIntelligent Voice ConferenceSpeech recognition-based meeting solutions[72]
Sales and RetailDigital Visit and SalesAI for FMCG optimization[72]
MediaIntelligent MediaAI for news production workflow[72]
Smart ParksPark Intelligent ManagementVisual AI for monitoring and decision-making[72]
Wildlife ProtectionAI Guardian of Endangered SpeciesDetects wildlife cybercrime[73]
WorkspaceOreateOne-stop AI workspace for documents, slides, images, video, and podcasts; multi-agent architecture[28]
No-code BuilderMiaoda 2.0No-code app builder; generated over 400,000 applications[28]
AI AgentGenFlowGeneral AI agent with 20 million users (November 2025); described by Baidu as the world's largest general-purpose agent[28]

Achievements and Recognition

  • China's designated "AI champion" by the government (2018)[1]
  • China's largest portfolio of AI-related patents and patent applications (December 31, 2024)[5]
  • Apollo Go received Gold in Driverless Vehicles category at 2025 Edison Awards[5]
  • PaddlePaddle recognized as leading Chinese deep learning framework[8]
  • Over 20% of Baidu search results are AI-generated (2025)[74]
  • ERNIE 4.5 claimed to outperform GPT-4.5 in multiple benchmarks at 1% of the cost[3]
  • ERNIE 5.0 ranked 1st among Chinese models and 8th globally on the LMArena leaderboard (January 2026)[80]
  • Apollo Go safety record of 10.14 million km per airbag deployment, significantly surpassing human driver averages[47]

Controversies

  • 2015 ImageNet incident: Baidu was accused of cheating on an AI image recognition test by submitting excessive entries, leading to a temporary ban from the competition[75]
  • 2024 AI model concerns: CEO Robin Li warned about "too many" AI models in China potentially leading to wasted resources[76]
  • 2025 doxxing scandal: Controversy involving Baidu's AI models led to calls for boycotts of its cloud services[77]

See Also

References

  1. "Baidu recognized as China's AI champion." Government of China AI Strategy Report, 2018.
  2. "Baidu chatbot 'Ernie' hits 200 million users." Yahoo Finance, April 2024.
  3. "Baidu launches ERNIE 4.5 and ERNIE X1." Baidu press release, March 2025.
  4. "Baidu Announces Fourth Quarter and Fiscal Year 2024 Results." Baidu Investor Relations, February 2025.
  5. "Baidu Q1 2025: AI Cloud and Robotaxi Fuel Growth." Baidu Investor Relations, May 2025.
  6. "Baidu Apollo Go robotaxi fleet reaches 1,000-vehicle milestone." CnEVPost, May 2025.
  7. "Apollo Go completes 2.2 million fully driverless rides in Q2 2025." Gasgoo Auto News, 2025.
  8. "Baidu's PaddlePaddle deep-learning platform fuels the rise of industrial AI." MIT Technology Review.
  9. "PaddlePaddle and ERNIE ecosystem reaches 21.85 million developers." Baidu, April 2025.
  10. "Baidu expands AI compute power with 3rd-gen Kunlun cluster deployment." Digitimes, 2025.
  11. "Baidu's Chip Unit Kunlunxin Files for Hong Kong IPO to Tap AI Investment Boom." Caixin Global, January 2, 2026. https://www.caixinglobal.com/2026-01-02/baidus-chip-unit-kunlunxin-files-for-hong-kong-ipo-to-tap-ai-investment-boom-102399601.html
  12. "Robin Li." Wikipedia.
  13. "Robin Li and the Algorithm that Built China's Internet Empire." Global Times Singapore.
  14. "Baidu founder Robin Li predicts 'exponential' boom in AI apps in 2025." South China Morning Post.
  15. "Baidu establishes Institute of Deep Learning." Baidu Research, 2013.
  16. "Andrew Ng joins Baidu as Chief Scientist." Baidu press release, 2014.
  17. "Qi Lu joins Baidu as COO." Baidu press release, January 2017.
  18. "Wang Haifeng appointed head of AI Group." Baidu corporate announcement, March 2017.
  19. "Baidu AI Group organizational structure." Baidu, 2017.
  20. "Wang Haifeng promoted to Senior Vice President." Baidu corporate announcement, May 2018.
  21. "Wang Haifeng promoted to CTO." Baidu corporate announcement, May 2019.
  22. "Baidu AIG division grows to 2,000+ employees." Baidu corporate report.
  23. "Baidu launches Ernie Bot." Global Times, March 2023.
  24. "Baidu discloses its ChatGPT rival now has 100M+ users." SiliconANGLE, December 2023.
  25. "Baidu Follows Alibaba's Steep LLM Price Cut With Two Free Ernie Bot Models." Yicai Global, May 2024.
  26. "Baidu unveils Ernie 4.0 Turbo." SiliconANGLE, June 2024.
  27. "Announcing the Open Source Release of the ERNIE 4.5 Model Family." ERNIE Blog, Baidu, June 30, 2025. https://ernie.baidu.com/blog/posts/ernie4.5/
  28. "Baidu Unveils ERNIE 5.0 and a Series of AI Applications at Baidu World 2025, Ramps Up Global Push." PR Newswire, November 13, 2025. https://www.prnewswire.com/news-releases/baidu-unveils-ernie-5-0-and-a-series-of-ai-applications-at-baidu-world-2025--ramps-up-global-push-302614531.html
  29. "Baidu Brain platform launched." Baidu, 2016.
  30. "Apollo autonomous driving platform." Apollo.auto.
  31. "ERNIE: Enhanced Representation through Knowledge Integration." Baidu Research, 2019.
  32. "ERNIE 3.0: Large-scale Knowledge Enhanced Pre-training for Language Understanding and Generation." arXiv:2107.02137.
  33. "Apollo Go receives fully driverless permits." Baidu press release, August 2022.
  34. "Baidu releases ERNIE 4.0." Baidu press release, October 2023.
  35. "RTA Signs MoU with Baidu's Apollo Go for autonomous taxi trials." Dubai Media Office, April 2025.
  36. "ERNIE 2.0: A Continual Pre-training Framework." Baidu Research, July 2019.
  37. "Introducing ERNIE 3.5: Baidu's Knowledge-Enhanced model." Baidu Research.
  38. "Baidu launches ERNIE 4.5 Turbo and X1 Turbo." Baidu press release, April 2025.
  39. "Baidu ERNIE Bot API calls reach 1.65 billion daily." Baidu corporate report, December 2024.
  40. "Baidu's AI bot has 300 million users, two months after reaching 200 million milestone." TechNode, July 2024.
  41. "Baidu Releases PaddlePaddle Framework 3.0." AIBase, 2025.
  42. "PaddlePaddle comprehensive toolsets." PaddlePaddle documentation.
  43. "Baidu Brain version 7.0 capabilities." Baidu, 2022.
  44. "Baidu autonomous driving early history." Baidu corporate history.
  45. "Baidu's Apollo Go robotaxi leads global autonomous driving with 17M+ orders." Car News China, November 2025.
  46. "China's Baidu says weekly robotaxi rides hit 250,000, same as Alphabet's Waymo." CNBC, November 2025.
  47. "Apollo Go safety record statistics." Baidu Apollo, 2025.
  48. "Baidu Apollo launches 6th-gen robotaxi with 60% lower cost." CnEVPost, May 2024.
  49. "Apollo Go seventh-generation robotaxi under $20,000." Industry reports, 2025.
  50. "How Baidu's Apollo Go Targets Global Robotaxi Expansion." Technology Magazine, 2025.
  51. "Baidu Qianfan GenAI platform accelerates enterprise adoption." Omdia, 2025.
  52. "Baidu Announces Third Quarter 2025 Results." Baidu Investor Relations.
  53. "DuerOS launched at Baidu World Conference." Baidu, 2015.
  54. "DuerOS conversational AI capabilities." Baidu DuerOS documentation.
  55. "DuerOS install base and voice queries." Baidu corporate report, June 2019.
  56. "DuerOS ecosystem partners." Baidu DuerOS.
  57. "Xiaodu smart device brand launched." Baidu, 2018.
  58. "Xiaodu Smart Display X10." Baidu Xiaodu product page.
  59. "XiaoduPods specifications." Baidu Xiaodu.
  60. "Xiaodu Qinghe AI tutor smartphone." Baidu, 2023.
  61. "Baidu Rolls Out Kunlun Roadmap, M100 and M300 AI Chips." TrendForce, November 2025.
  62. "Kunlun chips secure RMB 1 billion order from China Mobile." Digitimes, August 2025.
  63. "Baidu Research overview." Baidu Research website.
  64. "Silicon Valley AI Lab established." Baidu Research, 2014.
  65. "BIL and RAL labs established." Baidu Research, 2018.
  66. "National Engineering Laboratory of Deep Learning." China NDRC, 2017.
  67. "Alibaba holds wide lead over rivals in China's AI cloud market." South China Morning Post, 2025.
  68. "Who leads China's AI cloud race? Alibaba, ByteDance, and Baidu all stake a claim." KrASIA, 2025.
  69. "Wang Haifeng biography." Baidu corporate page.
  70. "Baidu Announces Fourth Quarter and Fiscal Year 2025 Results." PR Newswire, February 2026.
  71. "Baidu Announces Fourth Quarter and Fiscal Year 2024 Results." Baidu Investor Relations.
  72. "Baidu AI industry solutions." Baidu AI Cloud product pages.
  73. "AI Guardian of Endangered Species." Baidu AI for wildlife protection.
  74. "Over 20% of Baidu search results AI-generated." Industry reports, 2025.
  75. "Baidu ImageNet incident." MIT Technology Review, 2015.
  76. "Robin Li warns about too many AI models in China." Reuters, 2024.
  77. "Baidu 2025 doxxing controversy." Media reports, 2025.
  78. Robin Li, quoted in "Baidu Unveils ERNIE 5.0 and a Series of AI Applications at Baidu World 2025." PR Newswire, November 13, 2025. https://www.prnewswire.com/news-releases/baidu-unveils-ernie-5-0-and-a-series-of-ai-applications-at-baidu-world-2025--ramps-up-global-push-302614531.html
  79. "Baidu Announces Fourth Quarter and Fiscal Year 2025 Results." PR Newswire, February 26, 2026. https://www.prnewswire.com/news-releases/baidu-announces-fourth-quarter-and-fiscal-year-2025-results-302698026.html
  80. "Baidu's ERNIE 5.0 tops all Chinese AI models in LMArena with 2.4 trillion parameters." The Decoder, January 2026. https://the-decoder.com/baidus-ernie-5-0-tops-all-chinese-ai-models-in-lmarena-with-2-4-trillion-parameters/

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

4 revisions by 1 contributors · full history

Suggest edit