Wenxin 4.5 Turbo is released, but Li Yanhong says applications are the king.

Header image source: Baidu

In the year 2025, when the iteration speed of AI technology far exceeds expectations, a core proposition that developers and enterprises face is: how to avoid the application value being submerged by the technological wave when the model capabilities rapidly leap forward?

Baidu provided the answer at the Create Developer Conference on April 25. At the conference, the Wenxin Large Model 4.5 Turbo and X1 Turbo were officially unveiled, along with the simultaneous release of the Cangzhou OS content domain operating system. The general superintelligent app "Xiangxiang" has also been launched. While showcasing breakthroughs in technology, Baidu responded to the industry's eager expectations for the practical application of AI value with a comprehensive layout of scenario-based applications and the MCP (Model Context Protocol) ecosystem.

Li Yanhong stated to the developers on site, "As long as you find the right scenario, choose the right basic model, and sometimes learn a bit about model tuning methods, the applications developed on this basis will not be outdated." He emphasized, "Without applications, chips and models are of no value. There will be many models, but what will truly dominate the world in the future is applications; applications are the king."

Wenxin Large Model 4.5 Turbo: Multimodal, Low Cost, Strong Reasoning

Currently, the competition among AI large models has entered a heated stage, but issues such as single modality, reasoning hallucinations, and high operational costs still plague the industry. To address these pain points, Baidu has launched the Wenxin Large Model 4.5 Turbo and X1 Turbo, which feature multimodal capabilities, low costs, and strong reasoning as their core characteristics, showcasing a balance between technology and cost-effectiveness.

Wenxin 4.5 Turbo has further breakthroughs in multi-modal capabilities, with faster speeds and significantly lower prices—only 0.8 yuan for every million tokens input and 3.2 yuan output, which is as low as 40% compared to DeepSeek-V3. In multiple benchmark tests, 4.5 Turbo surpassed GPT-4o with an average score of 77.68 compared to 72.76, demonstrating its competitiveness. Meanwhile, X1 Turbo takes it a step further, based on the deep thinking optimization of 4.5 Turbo, with 1 yuan per million tokens input and 4 yuan output, just 25% of DeepSeek-R1, showing leading performance in scenarios such as Q&A, creative writing, logical reasoning, and tool invocation.

Li Yanhong stated at the scene: "Model iteration is fast, but the application value will not be overshadowed. As long as the right scenario is found, the right model is chosen, and fine-tuning is well utilized, AI applications can continue to create value." This perspective injects a boost of confidence to developers—the true value of large models lies in their deep integration with scenarios.

If the Wenxin large model is the "heart" of Baidu AI, then the newly released Cangzhou OS is the "nervous system" connecting content and applications. As a content operating system, Cangzhou OS is centered around Chatfile Plus and integrates three repositories: public knowledge base, private knowledge base, and memory base, as well as three tools: reader, editor, and player. It achieves the parsing and processing of multimodal content through the dynamic invocation of the large model.

Cangzhou OS | Photo source: Baidu

Relying on the Cangzhou OS, Baidu Wenku and Baidu Wangpan jointly launched the "AI Notes" feature. Users can generate structured notes and AI mind maps with one click while watching videos on Wangpan, and even create test questions based on video content, achieving a seamless connection between videos and notes. As of now, the number of paid users for Baidu Wenku's AI features has exceeded 40 million, with 97 million monthly active AI users; Baidu Wangpan's monthly active AI users exceed 80 million. This data not only reflects the productivity of the Cangzhou OS but also indicates the enormous potential of AI applications in the content field.

Another application of multimodal capabilities is Baidu's highly persuasive digital human. In the e-commerce live broadcast scenario, traditional digital humans tend to have boring copywriting and stiff movements, while Baidu's highly persuasive digital humans have achieved super-realistic voices, professional content, and flexible interaction through multi-modal large models. Its "script generation" function can adjust expressions, tone and actions in real time according to the lines, while the "AI brain" can dynamically schedule roles such as assistant broadcasters and field control according to the popularity of the live broadcast room, truly achieving "one person is a team". Baidu's Huiboxing platform has also launched a "one-click cloning" function, where users only need to upload a 2-minute video to generate an exclusive digital human, making "everyone can be an anchor" a reality.

Intelligent Agents and MCP: Igniting the Ecological Engine of AI Applications

As agents become a hot topic in the industry, Baidu has also launched the "Xinxiang" App to join the battlefield.

As a product centered around the "AI Task Completion Engine", Xinxiang achieves the breakdown, execution, and delivery of complex tasks through natural language interaction. Currently, Xinxiang covers over 200 tasks across ten major scenarios including in-depth research, intelligent charts, and game development, with plans to expand to over 100,000 in the future.

Xinxin supports the "multi-agent collaboration" mechanism. In health consultation scenarios, the system can schedule multiple "Doctor AI Avatars" for joint diagnosis; in legal services, it is the "Lawyer Think Tank" that collaborates on responses. This mechanism breaks through the limitations of traditional tool usage, providing a more efficient solution for professional scenarios. The free open access of Xinxin and the launch of the Android version (iOS version coming soon) also allow more users to experience the charm of AI agents.

The potential of multi-agent collaboration is also reflected in Baidu's no-code programming tool "Miaoda" launched last year. "Miaoda" can generate applications in a single sentence, and collaborate with multiple agents such as product managers, architects, and designers behind it, and call tool components such as cloud storage, databases, and map navigation. Using "Miaoda", a team of college students developed a marketing tool that used to take a week in just a few minutes, and the model call fee was only 5 yuan.

Li Yanhong stated: "There are 8 billion people in the world. When the technical threshold disappears, everyone can have the capabilities of a programmer." The open accessibility of "Miaoda" further lowers the development threshold for AI applications.

MCP (Model Context Protocol) is a new trend in the AI ecosystem that is reshaping the way developers interact with large models. Baidu announced at this conference that it is fully embracing MCP, launching the world's first e-commerce transaction MCP, search MCP, and other services, while optimizing the capabilities of the Wenxin large model in MCP task planning and scheduling.

At the same time, Baidu Intelligent Cloud's Qianfan platform has taken the lead in supporting MCP, allowing developers to create and publish MCP servers; Baidu Search has built an MCP server discovery platform that indexes high-quality services across the internet; Wenxin Kuai Ma has become the first intelligent coding assistant in China to support MCP servers. In addition, applications such as Baidu Wenku, Baidu Cloud, and Baidu Maps have fully opened MCP server services, while the MCP services launched by Baidu E-commerce cover functions such as product search, transactions, and parameter comparison, making it the first MCP service in China to support e-commerce transactions. This series of arrangements not only lowers the entry threshold for developers but also injects new momentum into the prosperity of the MCP ecosystem.

Baidu "Wenxin Cup" Entrepreneurship Competition Launch｜Image Source: Baidu

In addition to breakthroughs in technology and products, Baidu has also invested significantly in AI talent development and entrepreneurial support. The 5-year plan to cultivate 5 million AI talents has been completed ahead of schedule. During the conference, Li Yanhong announced that Baidu will train another 10 million AI talents in the next 5 years. Furthermore, the maximum investment amount for a single project in the third "Wenxin Cup" entrepreneurship competition has been raised to 70 million yuan, providing stronger support for entrepreneurs.

On the hardware level, Baidu has launched the country's first fully self-developed 30,000-card cluster, which can simultaneously support the training of multiple large models with hundreds of billions of parameters and allow 1,000 clients to fine-tune models with tens of billions of parameters. This cluster ensures the stability and energy efficiency of training tasks through high-performance networking and innovative cooling solutions.

The technology matrix released by Baidu reveals three major trends in AI development: multimodal capabilities becoming a foundational ability (Wenxin large model), intelligent agents reconstructing human-machine collaboration (Xinxiang App), and the MCP protocol unifying service interfaces (Baidu MCP ecosystem). As model capabilities break through thresholds, tool utilization achieves standardization, and computing costs continue to decline, a new era of AI applications is approaching. As Li Yanhong stated: "AI applications that can penetrate scenarios and bring real value are the true opportunities for developers."

"All these releases are aimed at allowing developers to focus on creating applications without worrying about model capabilities, model costs, or development tools and platforms, enabling them to devote themselves to making the best applications." Li Yanhong stated on site.

For the domestic AI industry, this series of actions by Baidu is not only a demonstration of technical strength but also a profound insight into scenario-based applications and ecological collaboration. In this transformation, Baidu is paving a "highway" for developers from model to application through technological innovation and ecological construction. With the continuous iteration of the Wenxin large model, Cangzhou OS, and Xingxiang, the "super productivity" of AI may be integrated into everyone's work and life even faster.