Artificial Intelligence (AI) literacy is gaining momentum in K-12 education, with active learning and small-group activities emerging as promising effective pedagogical strategies. However, teachers face challenges in monitoring and understanding small-group dynamics, which are crucial for enacting effective pedagogical tactics and providing timely support. Although it is becoming increasingly feasible to collect multimodal data (e.g., audio and video) in classrooms, translating these complex data streams into pedagogically relevant insights remains a challenge. The study is situated in an K-12 AI literacy project-based learning context and uses multimodal data streams collected from real classrooms, focusing on the process of social construction of knowledge emerging from small-group activities, including on-task versus off-task engagement and sharing versus beyond-sharing behaviors. We explored state-of-the-art multimodal machine learning models and demonstrated its advantage compared with unimodal approaches relying solely on features extracted from verbal messages. The results demonstrate the promise of multimodal models to support teaching analytics, even with a modest dataset collected in a noisy environment. We discuss the potential of this work to empower educators by leveraging complex multimodal data streams to support the process of promoting meaningful knowledge construction in AI literacy classrooms. This work contributes to the emerging body of research on the application of AI in real-world K-12 AI literacy settings and demonstrates how AI can serve not only as a subject of instruction but also as a tool to advance AI literacy education in an ecologically valid educational context.