.png)
April 27, 2025
Chinese tech giant Baidu unveiled two new language models at its Create 2025 developer conference in Wuhan: Ernie 4.5 Turbo and Ernie X1 Turbo. Both models are designed for processing text and images, as well as logical reasoning, and are offered at significantly lower prices than their predecessors.
According to Baidu, Ernie 4.5 Turbo operates faster and produces fewer errors in text generation compared to previous versions. The company claims its performance in text and image processing matches that of GPT-4.1 and surpasses GPT-4o. Benchmarks presented by Baidu show that Ernie 4.5 Turbo achieves a similar performance level to GPT-4.1 and GPT-4o in text and multimodal tasks. The price for Ernie 4.5 Turbo is 0.8 RMB (approximately 0.10 Euros) per million input text characters and 3.2 RMB (approximately 0.41 Euros) per million generated characters. Baidu states that this represents a cost reduction of 80 percent compared to the previous version.
With Ernie X1 Turbo, Baidu has also launched a turbo version of its reasoning model, first introduced in mid-March. Baidu reports that Ernie X1 Turbo outperforms competing models Deepseek-R1 and Deepseek-V3 in its performance. Here, too, Baidu substantiates its claims with benchmarks demonstrating that Ernie X1 Turbo delivers strong performance compared to competing models like Deepseek and OpenAI o1. The price for Ernie X1 Turbo is 1 RMB (approximately 0.13 Euros) per million input characters and 4 RMB (approximately 0.51 Euros) per million output characters. Baidu claims this is a quarter of the cost of Deepseek R1, which is already significantly cheaper than Western models. Baidu makes both models available for free within the Ernie Bot.
At Create 2025, which was held under the motto "Models Lead, APPs Rule," Baidu CEO Robin Li emphasized the importance of applications. He explained that AI models and chips are worthless without good apps. This perspective mirrors a recent comment by OpenAI's Go-to-Market Manager Adam Goldberg, who said that value creation for AI companies occurs across the entire chain.
Beyond language models, Baidu is expanding its focus to new AI applications. The company introduced the Huiboxing platform, which can generate digital avatars from short video clips. According to Baidu, these avatars will have a realistic appearance and a natural voice. In addition, Baidu presented Xinxiang, a multi-agent app developed for solving complex tasks. Baidu states that Xinxiang currently supports 200 task types, including knowledge analysis, travel planning, and office work. Baidu plans to expand this to over 100,000 task types and grant developers access. Xinxiang is available for Android, with an iOS version in testing. For technical integration, Xinxiang uses an extended version of Anthropic's Model Context Protocol (MCP), according to Baidu. Additionally, the company announced it would strengthen its commitment to developers through the "AI Open Initiative." This platform allows developers to market AI agents, mini-programs, and apps.
