The Zhitong Finance App learned that CICC released a research report stating that while maintaining Baidu Group-SW (09888)'s “outperforming industry” rating, revenue and non-standard net profit forecasts for 2024 and 2025 remain unchanged, with a target price of HK$146.9. Baidu held the Create2024 Baidu AI Developer Conference on April 16. At the conference, management announced the latest developments in the Wenxin Big Model and released new AI tools for developers.
The main views of CICC are as follows:
The scale of API calls is growing rapidly, and the underlying model continues to improve.
Currently, the number of Wenxin users has exceeded 200 million (announced at the end of December 23), the average number of daily API calls has exceeded 200 million (4Q23 financial report announced that it has exceeded 50 million), and performance and cost optimization have continued: compared to Wenxin 4.0 released in October 23, the efficiency of algorithm training has increased 4.1 times, the average weekly training efficiency has reached 98.8%, inference performance has increased 105 times, and inference costs continue to drop. In addition, the company innovatively introduced Comate, an intelligent code assistant in the technical architecture. Currently, the adoption rate of its code generation has reached 46%; in addition, the company released the Wenxin Big Model 4.0 tool version, which highlights documentation, data analysis, and code interpretation functions to facilitate developers' use.
Optimize the service architecture to improve the development efficiency and experience of AI native applications.
Based on continuous refinement of actual applications from internal and external customers, the company released a new intelligent computing operating system “Wan Yuan”:
Kernel layer: The 100 Ge AI heterogeneous computing platform is compatible with large-scale hybrid training of mainstream AI chips at home and abroad, and maintains low performance loss (100 calories and kcal hybrid training efficiency reaches 97% and 95%, respectively); in addition, the system's preset large models are comprehensive, including the flagship ERNIE 3.5 and 4.0, lightweight ERNIE Speed, Lite and Tiny, and the tripartite model. The company is expected to meet diverse needs in different scenarios.
Shell layer: ModelBuilder (model customization tool) provides model routing services, including model management, scheduling, and secondary development, and provides optimal model combinations under the MoE architecture (that is, mixing large and small models) to further reduce average inference costs. Currently, the number of ModelBuilder models has increased to 77, with built-in 47 data sets and 119 industry-level prompt templates.
Toolkit layer: supports rapid application development and release. AppBuilder (codeless programming tool) supports third-party APIs and custom components such as flight queries and paper searches, and also has cross-modal capabilities such as Wensheng Maps, and has now been applied in Baidu Library's smart comic functions; AgentBuilder (smart body development tool) currently supports the creation of 30,000 agents, covering more than 30 industries such as education, real estate, home, and business services.