Baidu AI

Baidu AI is an umbrella term for the artificial intelligence research, infrastructure, models, cloud services, and applications developed by Baidu. It is not a separately listed company or a standalone reportable financial segment. Baidu's 2025 annual report identifies two reportable segments, Baidu General Business and iQIYI. Within Baidu General Business, AI work is spread across the mobile ecosystem, AI Cloud, intelligent driving, marketing, and internal technology platforms.^[1][2]

The distinction matters because "Baidu AI" can refer to several different things. The historical Artificial Intelligence Group, or AIG, was an internal organization. "Baidu Core AI-powered Business" is a management reporting category introduced in 2025. ERNIE is a family of foundation models, while Ernie Bot and the newer ERNIE Assistant name consumer-facing applications. Qianfan is an enterprise model and agent platform. Apollo is an autonomous driving platform, and Apollo Go is the associated ride-hailing service.

For 2025, Baidu reported RMB40.0 billion of revenue from its Core AI-powered Business, up 48 percent from 2024. That category comprised AI Cloud Infrastructure, AI Applications, and AI-native Marketing Services. Baidu cautioned that these figures came from internal management accounts and were not audited.^[1][3] In the first quarter of 2026, the same category generated RMB13.6 billion and represented 52 percent of Baidu General Business revenue.^[4] Neither figure is revenue for ERNIE alone, and neither includes every activity that could reasonably be called AI.

Business scope

Baidu describes its strategy as a stack extending from computing infrastructure to applications. The main layers are:

Scroll sideways for more →

These layers share technology, but they are not one product. A customer can use Qianfan without downloading ERNIE weights, deploy a third-party model on Baidu infrastructure, or build with PaddlePaddle outside Baidu AI Cloud. Apollo Go also has its own vehicles, operating systems, local approvals, fleet operations, and partners.

Baidu's public company overview places Ernie Bot in the mobile ecosystem, AI Cloud Infrastructure and AI Applications in the cloud business, and Apollo Go and DuerOS in Intelligent Driving and Other Growth Initiatives.^[2] DuerOS is a voice assistant used in Xiaodu devices and other products. Historical installation and voice-query counts from 2019 do not describe the current user base, so they are not carried forward here.

History

Baidu was founded by Robin Li and Eric Xu in 2000 as an internet search company. Li studied information science at Peking University before earning a master's degree in computer science in the United States.^[25] Baidu's use of machine learning grew from search ranking, advertising, language processing, and recommendation. A 2017 Baidu filing says the company established its Institute of Deep Learning in January 2013, opened the Silicon Valley AI Lab in May 2014, and later operated research labs in deep learning, big data, Silicon Valley AI, and augmented reality.^[5]

In May 2014, Baidu appointed Andrew Ng, a co-founder of Google Brain, as chief scientist and head of Baidu Research. The appointment accompanied a new Silicon Valley laboratory focused on fields including image and speech technology.^[6] Ng left in 2017.

Baidu open-sourced PaddlePaddle in September 2016. In 2017 it consolidated several technology teams into an Artificial Intelligence Group led by Haifeng Wang.^[26] Baidu also announced Project Apollo in April 2017 as an open platform for autonomous-driving software and data.^[5] AIG remains useful historical context, but current filings do not present "Baidu AI" or AIG as a reportable segment. Baidu promoted Wang to senior vice president and general manager of AIG in 2018.^[7]

The ERNIE line began in 2019 with a language-representation paper. Baidu presented Ernie Bot in March 2023 during the surge in generative AI products following ChatGPT. The service opened to the general public on August 31, 2023, after China's interim rules for public generative-AI services took effect.^[8][9] Baidu made Ernie Bot free for consumers in April 2025.^[1]

By 2025 and 2026, Baidu had shifted its public reporting toward commercialization. It began disclosing AI Cloud Infrastructure, AI Applications, and AI-native Marketing Services as a management-defined revenue category, proposed a separate listing for its Kunlunxin chip subsidiary, expanded Apollo Go outside mainland China, and released the ERNIE 4.5, 5.0, and 5.1 model generations.^[1][3][4]

ERNIE models and applications

Model development

ERNIE expands to Enhanced Representation through Knowledge Integration. The 2019 paper introduced entity-level and phrase-level masking rather than masking only individual tokens. It reported results on Chinese natural-language-processing tasks, but it was a Baidu-authored preprint and its benchmark findings should be read in that context.^[8]

ERNIE 3.0, described in a 2021 Baidu paper, combined autoencoding and autoregressive pre-training. The authors reported a 10-billion-parameter model trained on a 4 TB corpus of text and a large knowledge graph.^[10] Later ERNIE releases are separate checkpoints and product families, so their parameter counts, context windows, licenses, modalities, and access terms should not be blended into one timeless specification.

In October 2023, Baidu introduced ERNIE 4.0 and said it was comparable with GPT-4 in comprehension, reasoning, memory, and generation. Independent reporting correctly treated parity as Baidu's claim rather than an established cross-model result.^[11] Model comparisons depend on the checkpoint, prompt set, scoring method, language mix, tool access, and evaluation date.

Baidu released ten ERNIE 4.5 variants in June 2025. Its repository lists language and vision-language mixture-of-experts models plus a small dense model. The largest listed vision-language variant has 424 billion total parameters and 47 billion active parameters. Baidu published model weights and inference code under Apache 2.0 through GitHub, AI Studio, and Hugging Face.^[12] This open availability applies to those artifacts, not automatically to every ERNIE service or later release.

The distinction between open weights and fully reproducible open source is especially relevant. In a January 2026 interview, Robin Li said "open weight" was the more exact description because the training data needed to reproduce the models were not public.^[13] Licenses, code, data documentation, evaluation assets, and hosted-service terms therefore need to be checked separately.

ERNIE 5.0's February 2026 technical report describes a 2.4-trillion-parameter mixture-of-experts model that processes text, images, video, and audio in one autoregressive framework. Baidu says fewer than 3 percent of parameters are activated for a token.^[14] These architecture details come from Baidu's own technical report, and its performance tables are vendor evaluations rather than independent certification.

Baidu launched ERNIE 5.1 in May 2026 as a smaller successor focused on text, reasoning, search, and agent tasks.^[4][15] Baidu's leaderboard and cost statements are point-in-time company claims. The available model in a consumer application can also differ from the model offered through Qianfan or from a downloadable ERNIE 4.5 checkpoint.

Consumer and productivity applications

Ernie Bot began as a chatbot and model interface, but Baidu has increasingly integrated ERNIE into search, document productivity tools, cloud storage, digital humans, code generation, and agents. Its current company description identifies Baidu Wenku and Baidu Drive as AI Applications. The first-quarter 2026 release also lists DuMate, Miaoda 3.0, Famou Agent 2.0, and GenFlow 4.0.^[2][4] These are separate products whose availability and capabilities can change independently.

Baidu reported 200 million registered Ernie Bot users in April 2024.^[16] A registered-user count is not the same as monthly active users, daily active users, subscribers, or API customers. Baidu later reported 202 million monthly active users for ERNIE Assistant in December 2025, a different product label and metric.^[3] Those values should not be combined into a continuous "user growth" table without defining the service and denominator.

Applications based on a large language model can produce inaccurate, incomplete, or policy-constrained answers. Access method also matters: the consumer assistant, a Qianfan endpoint, and a self-hosted open-weight model can have different system instructions, retrieval sources, moderation, retention terms, and update schedules.

PaddlePaddle

PaddlePaddle is Baidu's open-source framework for training and deploying neural networks. Baidu's 2017 filing records its public release in September 2016 for tasks including natural-language processing, image recognition, and speech recognition.^[5] Baidu's 2025 annual report says PaddlePaddle 3.0 was released in April 2025 and was designed to support large-model training and inference.^[1]

The ecosystem includes libraries for language, optical character recognition, vision, speech, recommendation, and distributed training. It occupies a similar framework layer to TensorFlow and Meta AI's PyTorch, but compatibility and performance depend on the model, accelerator, operators, distributed setup, and deployment target. Claims that one framework "leads China" need a named market definition and dated methodology.

PaddlePaddle is also the training and deployment base described for ERNIE 4.5 and 5.0. That relationship does not mean every model runnable in PaddlePaddle was made by Baidu, or that using PaddlePaddle requires Baidu AI Cloud.

Baidu AI Cloud and Qianfan

Baidu AI Cloud sells compute, storage, model training and inference, and higher-level applications. Qianfan is its managed model and agent-development platform. Current international documentation describes a unified API for ERNIE and third-party open models, an agent engine, tools and Model Context Protocol integrations, and OpenAI-compatible REST interfaces.^[17]

Qianfan is therefore broader than an ERNIE API. Its June 25, 2026 international model list included ERNIE and third-party families such as DeepSeek and Kimi, with model-specific context limits and rate limits.^[18] The Chinese and international services can have different endpoints, models, prices, quotas, and terms. A current deployment decision should use the live regional documentation rather than a historical inventory.

In the first quarter of 2026, Baidu reported RMB8.8 billion of AI Cloud Infrastructure revenue, up 79 percent year over year, and RMB2.5 billion of AI Applications revenue, approximately flat.^[4] Baidu defines the first category to include AI infrastructure and platform services. It does not disclose how much of that revenue came from Qianfan, ERNIE inference, custom projects, or any one customer group.

Kunlunxin and computing infrastructure

Kunlunxin develops AI chips and integrated hardware-software systems. It grew from Baidu's internal accelerator work into a non-wholly owned subsidiary. Baidu also uses third-party accelerators and heterogeneous computing infrastructure, so Kunlunxin should not be described as the sole hardware basis for every Baidu model or cloud service.^[1]

On January 1, 2026, Kunlunxin confidentially submitted an application for a Hong Kong Stock Exchange listing. Baidu called the transaction a proposed spin-off and said Kunlunxin was expected to remain a subsidiary after completion. The company also warned that details were not final and there was no assurance the listing would occur.^[19] A filed application is not a completed initial public offering.

Detailed chip roadmaps and performance figures in vendor announcements are difficult to compare with products from other suppliers. Process node, numeric format, memory, interconnect, compiler support, power envelope, system size, and tested workload all affect a comparison. A TOPS or FLOPS figure alone does not establish training speed, inference cost, or availability.

Apollo and Apollo Go

Apollo is Baidu's autonomous-driving software platform. The project was announced in 2017, and its public repository contains perception, localization, prediction, planning, control, mapping, simulation, and developer tools.^[5][20] Apollo Go is different: it is an operating ride-hailing service that combines vehicles, the driving system, fleet supervision, maps, customer software, and city-specific permits.

Baidu reported 3.2 million fully driverless operational rides in the first quarter of 2026, more than 22 million cumulative public rides by April, and 330 million autonomous kilometers by May, including 220 million fully driverless kilometers. It also described a 27-city global footprint.^[4] These are company-reported operational figures. A "footprint" can include testing, pilots, partnerships, and commercial service, and does not mean identical public availability in every city.

International deployments use local partners and approvals. In February 2026, Baidu and Uber announced that Apollo Go vehicles would be available through Uber in selected parts of Dubai, with fleet management by a third-party operator and expansion subject to operating results and regulatory approval.^[21] Announced testing in another country is not the same as a launched public service.

Direct comparisons with Waymo require consistent definitions. Weekly rides, cumulative rides, paid rides, public rides, operational rides, miles, kilometers, service areas, and intervention reporting are not interchangeable. The baseline's claim that Apollo Go "matched Waymo" used differently scoped company disclosures and is omitted.

On March 31, 2026, more than 100 Apollo Go vehicles stopped in Wuhan after what police described as a system malfunction. Some passengers were stranded in moving traffic; no injuries were reported. Baidu did not immediately comment to the Associated Press.^[22] The incident shows why aggregate mileage and ride counts should be considered alongside outage response, remote assistance, incident reporting, local regulation, and independent safety evidence.

Baidu describes its safety record favorably, but no common public dataset establishes a universal Apollo Go-versus-human or Apollo Go-versus-competitor safety ratio. Exposure conditions, crash definitions, reporting thresholds, road types, weather, and human supervision must be comparable before drawing that conclusion.

Research and leadership

Robin Li is Baidu's co-founder, chairman, and chief executive officer. Baidu's 2025 annual report also names Haifeng Wang as chief technology officer and an executive officer.^[1] Wang joined Baidu in 2010 and led the historical AIG. The article does not reproduce a long executive roster because group-level roles and internal business-group leadership can change.

Baidu Research has worked in natural-language processing, computer vision, speech, knowledge graphs, recommendation, robotics, and autonomous driving. Research papers and code releases are useful evidence for architecture and reproducibility, but author-reported benchmark results should not be treated as independent product tests.

One important historical qualification is the 2015 ImageNet test-server incident. Challenge organizers reported that a Baidu team used at least 30 accounts to make at least 200 submissions between November 2014 and May 2015, far above the two-per-week limit. They said the resulting scores were not directly comparable and asked the team to refrain from submissions for 12 months. The team apologized in the organizers' statement.^[23] This was a rules and evaluation-integrity failure; it does not invalidate unrelated Baidu research, but it is relevant when discussing benchmark governance.

The broader deep learning shift associated with the 2012 ImageNet breakthrough and researchers such as Geoffrey Hinton helped motivate large corporate AI laboratories. It does not establish that one company originated the field or that modern model quality can be inferred from old image-classification results.

Financial performance

Baidu reported total 2025 revenue of RMB129.1 billion, down 3 percent year over year. Its management-defined Core AI-powered Business generated RMB40.0 billion, or 39 percent of Baidu General Business revenue. The components were RMB19.8 billion from AI Cloud Infrastructure, RMB10.2 billion from AI Applications, and RMB9.8 billion from AI-native Marketing Services.^[3]

In the first quarter of 2026, those three components were RMB8.8 billion, RMB2.5 billion, and RMB2.3 billion, respectively. Together they were RMB13.6 billion, up 49 percent year over year.^[4] Baidu General Business revenue was RMB26.0 billion, and total Baidu revenue, including iQIYI, was RMB32.1 billion.^[4]

The reporting boundaries impose several limits:

• The AI-powered figures are unaudited internal management data even though Baidu's consolidated annual financial statements are audited.
• The category definition was introduced in 2025 and may change, limiting comparisons with earlier periods or other companies.
• AI Cloud Infrastructure is not the same as total cloud revenue, and AI Applications is not the same as Ernie Bot revenue.
• Apollo Go operational metrics are disclosed separately; the releases do not assign Apollo Go a standalone revenue figure.
• A growth rate does not reveal profitability, customer concentration, contract length, or the cost of computing capacity.

Competition and market interpretation

Baidu competes with large Chinese technology companies and specialist model, cloud, chip, and autonomous-driving firms. Its annual report names competition across foundation models, generative AI, cloud services, search, marketing, and robotaxis.^[1] Relevant products include Alibaba Cloud and Qwen, ByteDance's Doubao, Tencent, Huawei Cloud, and independent model developers.

Baidu's combination of models, framework, cloud, chips, search, and robotaxis is broad, but it is not unique in owning multiple stack layers. The baseline statement that no rival combined comparable capabilities was unsupported. Product leadership also cannot be inferred from one company-wide label: cloud revenue, model API traffic, chatbot use, open-weight downloads, benchmark scores, and robotaxi rides measure different markets.

There is no single "Baidu AI market share." Analyst estimates for Chinese public cloud, AI cloud, hosted-model tokens, consumer assistants, and robotaxis use different denominators and often disagree. A reliable comparison should name the category, geography, period, currency, inclusion rules, and source methodology. The candidate therefore omits the baseline's market-share table.

The same caution applies to frontier models. ERNIE, DeepSeek, Qwen, and models from Google or other developers can be compared on a defined task, but a leaderboard snapshot does not establish general superiority. Pricing claims also change with caching, batch discounts, context length, input-output mix, region, and hosted versus self-managed deployment.

Regulation and responsible interpretation

China's Interim Measures for the Management of Generative Artificial Intelligence Services took effect on August 15, 2023. The rules cover public-facing services and include obligations involving information security, personal data, service governance, and protection of minors.^[9][24] Ernie Bot's public opening later that month followed the new approval and security-review environment.^[9]

Regulation affects consumer assistants, cloud services, and autonomous vehicles differently. A public chatbot, an internal enterprise model, an open-weight download, and a robotaxi pilot can face different approvals and operational duties. International deployments add local transport, privacy, mapping, data-transfer, and product-liability rules.

Readers should also separate four kinds of evidence:

• Audited company financials describe Baidu's consolidated results, not every AI product.
• Unaudited management metrics describe Baidu's own internal categories and operational counts.
• Technical papers and repositories document architectures and artifacts, often from the developer.
• Independent reporting and research can test or contextualize claims but may still use incomplete public data.

This distinction is more useful than a single ranking. It also prevents vendor benchmarks, registered-user totals, model downloads, ride counts, and proposed product roadmaps from being presented as equivalent proof of technical quality or commercial success.

See also

• Transformers

References

1. ^Baidu, Inc., "Annual Report on Form 20-F for the fiscal year ended December 31, 2025," March 17, 2026. sec.gov/...d38065d20f
2. ^Baidu, Inc., "Company Overview," accessed July 28, 2026. ir.baidu.com/company-overview
3. ^Baidu, Inc., "Baidu Announces Fourth Quarter and Fiscal Year 2025 Results," February 26, 2026. ir.baidu.com/...arter-and-fiscal-year-2025-results
4. ^Baidu, Inc., "Baidu Announces First Quarter 2026 Results," May 18, 2026. ir.baidu.com/...nounces-first-quarter-2026-results
5. ^Baidu, Inc., "Prospectus Supplement," June 28, 2017. sec.gov/...d377819d424b2
6. ^Baidu, Inc., "Baidu Opens Silicon Valley Lab, Appoints Andrew Ng as Head of Baidu Research," May 16, 2014. ir.baidu.com/...-lab-appoints-andrew-ng-head-baidu
7. ^Baidu, Inc., "Qi Lu to Transition into New Role at Baidu, Baidu Promotes Haifeng Wang to Senior Vice President," May 18, 2018. ir.baidu.com/...aidu-baidu-promotes-haifeng-wang-0
8. ^Yu Sun, Shuohuan Wang, Yukun Li, et al., "ERNIE: Enhanced Representation through Knowledge Integration," arXiv:1904.09223, 2019. arxiv.org/...1904.09223
9. ^Associated Press, "China's Baidu makes AI chatbot Ernie Bot publicly available," August 31, 2023. apnews.com/...627bd09608816847907d41f44da235d9
10. ^Yu Sun, Shuohuan Wang, Shikun Feng, et al., "ERNIE 3.0: Large-scale Knowledge Enhanced Pre-training for Language Understanding and Generation," arXiv:2107.02137, 2021. arxiv.org/...2107.02137
11. ^Associated Press, "Chinese search engine company Baidu unveils Ernie 4.0 AI model, claims that it rivals GPT-4," October 17, 2023. apnews.com/...e7fa36ad28215db42f720a3ac028e2c4
12. ^Baidu and PaddlePaddle, "ERNIE 4.5 and ERNIEKit," GitHub repository, accessed July 28, 2026. github.com/...ERNIE
13. ^Charlie Campbell, "'We're Not That Far Behind.' Baidu's Robin Li on China's Push to Diffuse AI Throughout Society," Time, January 25, 2026. time.com/...robin-li-baidu-interview
14. ^Haifeng Wang et al., "ERNIE 5.0 Technical Report," February 6, 2026. arxiv.org/...2602.04705
15. ^Baidu, "ERNIE 5.1 Officially Released," May 9, 2026. ernie.baidu.com/...ernie-5.1-0508-release
16. ^Reuters, "Baidu says AI chatbot Ernie Bot has amassed 200 million users," April 16, 2024. finance.yahoo.com/...ys-ai-chatbot-ernie-020254560
17. ^Baidu AI Cloud, "Baidu Qianfan," accessed July 28, 2026. intl.cloud.baidu.com/...qianfan
18. ^Baidu AI Cloud, "Qianfan Model List," updated June 25, 2026. intl.cloud.baidu.com/...7m95lyy43-intl-en
19. ^Baidu, Inc., "Baidu Announces Proposed Spin-off and Separate Listing of Kunlunxin," January 1, 2026. ir.baidu.com/...pin-and-separate-listing-kunlunxin
20. ^ApolloAuto, "Apollo: An open autonomous driving platform," GitHub repository, accessed July 28, 2026. github.com/...apollo
21. ^Baidu and Uber, "Baidu and Uber Partner to Bring Apollo Go Autonomous Ride-hailing to Dubai," February 10, 2026. investor.uber.com/...default
22. ^Associated Press, "Passengers stranded in moving traffic after robotaxi outage in China's Wuhan," April 1, 2026. apnews.com/...ec3dcd026dfa1a19eb79467d5211f01d
23. ^ImageNet Large Scale Visual Recognition Challenge organizers, "ILSVRC announcement," June 2, 2015. image-net.org/...announcement-June-2-2015
24. ^State Council of the People's Republic of China, "China moves to support generative AI, regulate applications," July 13, 2023. english.www.gov.cn/...t_WS64aff5b3c6d0868f4e8ddc01
25. ^Baidu, Inc., "Robin Li," accessed July 28, 2026. ir.baidu.com/...robin-li
26. ^Shanghai Daily, "Baidu AI expert leaves," March 23, 2017. archive.shine.cn/...shdaily.shtml