Om AI Launches Physical AI Initiative at WAIC 2026, Industry Reaches Consensus on Native End-Side Architecture

On July 19, Om AI Lianhui hosted the "End-Side Streaming Multimodal Models Empowering AI Hardware Boom" forum at the Shanghai World Expo Center during WAIC 2026 — the only forum at this year's conference dedicated to end-side multimodal models and AI hardware deployment. At the event, the company formally launched the "End-Side Multimodal Base Model and AI Hardware Collaborative Development Initiative" and released the VLX developer community, marking a shift from isolated technical exploration to industry-wide collaboration.

Current mainstream end-side AI solutions predominantly rely on a "cloud pre-training plus terminal pruning" approach. However, physical-world scenarios impose rigid requirements for low-latency real-time response, offline independent operation, and local data privacy, exposing the limitations of cloud-only methods. The forum brought together experts across the full industry chain — from chip architecture and computing power to data security and terminal manufacturing — who reached a clear consensus that native end-side architecture is more suitable for commercial-scale physical AI deployment.

Om AI CEO and Chief Scientist Zhao Tiancheng noted that the real-time, dynamic, and high-frequency interactive nature of the physical world means end-side AI cannot simply be a stripped-down version of cloud models — its underlying architecture must be reconstructed for physical scenarios. Multiple attending scholars further confirmed that the core bottleneck for embodied AI at scale lies in real-time perception and dynamic adaptation at the terminal, and that local processing advantages have shifted from being a constraint to becoming a competitive driver.

As one of the first companies in China to focus on end-side native architecture, Om AI has developed the VLX series of end-side streaming multimodal models specifically for physical-world applications, targeting issues of high latency, poor adaptation, and weak real-time performance. The VLX series, previously available for online trials, made its offline public debut at WAIC 2026, drawing significant industry attention.

The initiative proposes three collaborative directions: establishing native end-side technical standards for evaluation systems, interface specifications, and security frameworks; jointly developing integrated hardware-software solutions for model-chip coordination and edge-cloud boundaries; and expanding physical AI industry scenarios through data feedback mechanisms and open ecosystems.

The VLX developer community launched alongside the initiative provides the technical foundation for implementation, offering global hardware manufacturers, system integrators, and AI developers full access to VLX capabilities with standardized APIs and comprehensive SDK support. Two in-depth industry dialogues at the forum validated the value of end-side technology from both consumer and industrial perspectives, with participants from low-altitude economy and maritime operations confirming that end-side AI has become an industry necessity rather than an optional upgrade. "The physical world is not a simple extension of cloud-based digital AI," Zhao Tiancheng said. "End-side intelligence needs its own architectural approach, evaluation system, security framework, and ecosystem."