AI

The history of Baidu’s Ernie Bot and the 2026 AI market

Baidu's Ernie Bot grew to 300 million users by June 2024 and evolved into the efficient Ernie 5.1 model. The 2026 AI market features intense competition between Baidu, Doubao, and Qwen as OpenAI faces restrictions in 34 markets.

The history of Baidu's Ernie Bot and the 2026 AI market

The Launch and Rapid Adoption of Ernie Bot

Baidu launched Ernie Bot for invited testing on March 16, 2023. The company stock dropped 10 percent that day because the live demonstration used prerecorded footage. However, the stock gained 14 percent the following day after analysts from Citigroup and Bank of America provided positive preliminary reviews. Before the public release, the ERNIE series had been in development since 2019. Following regulatory approval, the service released to the general public on August 31, 2023. The user base grew to 100 million by December 2023 and reached 300 million users by June 2024. In January 2024, Ernie Bot was integrated into Samsung’s Galaxy S24 lineup. Baidu denied reports in January 2024 that a university research lab linked to the People’s Liberation Army had tested the bot for military response scenarios. In September 2024, Baidu rebranded the chatbot to "Wenxiaoyan" to act as a search assistant. The chatbot follows strict censorship rules and refuses to answer questions regarding the leadership of Xi Jinping or the 1989 Tiananmen Square protests. Ernie 4.5 was open sourced under Apache 2.0 in June 2025, which allowed for widespread deployment of its variants.

Model Evolution and Efficiency Gains

Baidu expanded its model capabilities through several iterations. The company released Ernie 4.0 in October 2023, which offered improved performance over previous versions and information updated to April 2023. The initial model used 10 billion parameters and a 4-terabyte corpus of plain text and knowledge graph data. In March 2025, Baidu announced Ernie 4.5 and the reasoning model Ernie X1. In April 2025, the company released Wenxin 4.5 Turbo and X1 Turbo to provide faster response times and lower operational costs. You know the fundamentals of LLM scaling, so the shift toward parameter efficiency in 2026 is a notable strategic pivot. Baidu’s developers used a technique called Once-for-All elastic training to build a super network that simultaneously optimizes various sub-networks across depth, width, and sparsity during the entire massive pre-training process. Ernie 5.1, released in May 2026, uses roughly 800 billion total parameters, which is about one third of the 2.4 trillion parameters found in Ernie 5.0.

Ernie 5.1 Metric Value
Total Parameters 800 billion
Global Search Rank #4 (LMArena)
Training Efficiency 6% of 5.0 compute
Professional Plan 59.90 yuan/month

The professional plan provides access to Ernie 4.0 for 59.90 yuan per month, while Ernie 3.5 remains free. Ernie 5.1 ranks #4 globally on the LMArena Search Arena, though it trails Gemini 3.1 Pro in spreadsheet automation and advanced instruction following.

The 2026 AI Search Market

The Chinese AI market reached a competitive peak in 2026. Baidu changed its traditional keyword-based platform to an AI-driven assistant. In August 2026, Baidu reported a 19 percent year-over-year fall in online marketing revenue because of a deliberate delay in AI-search monetization. In May 2026, native AI apps in China reached 499 million monthly active users, with Doubao at 382 million and Ernie Bot at 220 million according to Statista. This competition involves a "red-envelope war" among giants like Alibaba, Tencent, and ByteDance to capture users. Doubao reached 382 million monthly active users by May 2026, while Alibaba’s Qwen app reached 167 million. The restriction of OpenAI’s advanced features in 34 markets helped Chinese developers use domestic models without competing against the most advanced global algorithms. Baidu targets the massive 700 million monthly active user base on the Baidu App to avoid the need for separate downloads. On mobile devices, Baidu maintains 73.49 percent market share, although Bing leads on desktop with 39.37 percent. Will the convergence of the "Chip – Cloud – Model – Agent" stack provide a permanent advantage against rivals like Doubao and Qwen?