Gasgoo Munich-The 2026 World Artificial Intelligence Conference opened today at the Shanghai World Expo Exhibition and Convention Center. ModelBest used the stage to showcase the commercial progress of its edge-side models across automotive, robotics, consumer electronics, and aerospace, while also unveiling its MiniCPM-Robot series for embodied AI.
Image Credit: ModelBest
On the show floor, a venue guide Agent developed jointly with Leju Robotics is running entirely offline on the edge. It supports real-time Q&A based on a local knowledge base and navigates obstacles autonomously. A separate patrol system identifies parking violations and road anomalies, processing all sensitive data locally. In the smart cockpit space, the SuperMate product is already shipping in mass-production models like the Changan Mazda EZ-60 and Geely Galaxy M9. Leveraging vision, reasoning, and memory, the system shifts the in-car experience from passive response to proactive service.
Image Credit: ModelBest
On the consumer electronics front, MiniCPM edge models are set to come pre-installed on several Samsung flagship devices. ModelBest also demonstrated its AgentVerse multi-agent system, covering scenarios in education, healthcare, and high-end manufacturing. The company's AI-native hardware, the "Pinecone Pie," serves as an edge-side development board designed to lower the barrier for AI application development. Additionally, a "Large Model × Space Computing" solution developed with Yiwu Aerospace is being used for on-orbit intelligent warnings and remote sensing for agriculture, forestry, and fisheries.
Image Credit: ModelBest
Technologically, ModelBest reiterated its "Density Law" — the principle that the parameter count required for a specific level of intelligence halves every 3.5 months. After two and a half years of iteration, the MiniCPM series has built a comprehensive edge-side capability stack spanning text, vision, omni-modal, and voice domains. Notably, MiniCPM-o 4.5 supports offline, real-time, omni-modal dialogue. Furthermore, the proprietary ForgeTrain Agent system outperformed Google’s 270M-parameter Gemma 3 model using just 130M parameters on Ascend chips, validating the efficiency of the domestic training loop.
Image Credit: ModelBest
According to ModelBest, cumulative downloads for its open-source MiniCPM series have surpassed 38 million across GitHub and Hugging Face. From an unconventional start to edge intelligence becoming a central theme at WAIC, ModelBest has established a complete commercial loop from model to scenario.
