Baidu
Why it matters
Deep Dive
Baidu spent two decades as China's dominant search engine, but it has invested in AI research since the early 2010s and now operates a stack few companies match: Kunlun accelerators at the bottom, the PaddlePaddle framework, the Ernie model line, and services like search, maps, cloud, and robotaxis on top. That position let it move fast when generative AI broke — Ernie Bot launched in March 2023, only months after OpenAI's ChatGPT, making it the first high-profile Chinese response. Since then the company has shipped new Ernie generations at a steady clip, pivoted from closed to open licensing, and fought for share in a domestic market where model prices have collapsed.
From ERNIE to Ernie Bot
Baidu's model work predates the chatbot era. The original ERNIE (Enhanced Representation through Knowledge Integration) appeared in 2019 as a pre-trained language model that fused knowledge-graph information into standard masked-language-modeling pretraining, and it scored at the top of Chinese and English NLP benchmarks of its day. That head start mattered when ChatGPT reset the industry: Ernie Bot (Wenxin Yiyan) launched in March 2023 and was among the first generative services approved for the Chinese public when regulators began licensing them later that year. Ernie 4.0 followed in late 2023 with better reasoning, memory, and multimodal understanding, and Baidu spent 2024 pushing Ernie into search, maps, docs, and its cloud API.
Ernie 4.5, X1, and the Open-Source Turn
The 2025 generation is the one practitioners should know. Ernie 4.5 is a natively multimodal family built on a mixture-of-experts architecture, while Ernie X1 is a reasoning model that works through extended chains of thought before answering and was priced aggressively against its Chinese rivals. The bigger surprise was the license: after years of publicly arguing that closed models would stay ahead, Baidu reversed course in mid-2025 and released the Ernie 4.5 family under Apache 2.0, from a 0.3-billion-parameter dense model up to a 424-billion-parameter MoE flagship with 47 billion active parameters.
The logic tracked what DeepSeek had demonstrated months earlier: open weights buy developer mindshare, enterprise pilots, and international reach that API-only models struggle to get, especially for a company whose consumer chatbot was losing ground at home. Releasing weights also turns the model into a distribution channel for Baidu's cloud and chips, since every self-hosted deployment is a candidate for managed hosting later. The tradeoff is real, though: once weights are public, competitors can fine-tune and distill them freely, which is exactly the dynamic that eroded Baidu's early chatbot lead.
Robotaxis and Silicon
Baidu's AI ambitions reach well past chatbots. Apollo Go, its robotaxi service, runs fully driverless commercial rides in several Chinese cities — Wuhan hosts the largest fleet — and has completed millions of paid passenger trips while beginning to expand into overseas markets. On the hardware side, the company designs the Kunlun line of AI accelerators and deploys them at scale in its own data centers for training and inference, a strategic hedge against export controls on high-end NVIDIA GPUs. Add PaddlePaddle, the open-source deep-learning framework Baidu has maintained since 2016, and the result is a chips-to-apps stack matched by only a handful of companies worldwide.
Not Just "China's Google"
A common misconception is that Baidu is a search company dabbling in AI — or, from the other direction, a fully realized AI powerhouse. The truth sits between. Baidu opened a dedicated deep-learning lab in 2013, hired Andrew Ng as chief scientist in 2014, and ran one of China's earliest large-scale NLP programs, so the research depth is genuine and long-standing. But search advertising still pays the bills, and AI revenue leans heavily on a cloud unit fighting a brutal price war, which means Ernie's progress is paced by what the core business can carry. Treat Baidu like an infrastructure incumbent making a long bet — well funded and durable, but not immune to the economics around it.
The China Landscape and the Price War
Baidu operates in the world's most crowded foundation-model market. Alibaba's Qwen series leads open-model adoption, ByteDance's Doubao dominates consumer chatbot usage in China, Tencent ships its Hunyuan models across WeChat and its games empire, and DeepSeek keeps resetting everyone's price expectations. That pressure is why Baidu made Ernie Bot free for consumers in 2025 and why Chinese API prices fell by more than 90 percent in some tiers during the 2024–2025 price war. For buyers the result is a market where capable models cost a fraction of Western equivalents — and where Ernie is one credible option among several rather than the default choice.