From September 22 to 24, the Apsara Conference 2026 was held at the Hangzhou International Expo Center. Under the theme of a dual-chip architecture accelerating edge-AI deployment, Rockchip presented its latest advances at booth 2C07 in Hall 2, including a VLA (vision-language-action) robot policy model, a multi-chip cascading solution and the VioLoop personal AI assistant, covering the full chain from silicon and model deployment to real-world applications.
Around the goal of keeping inference local, an offline on-device AI office agent jointly developed with Yici Wanglian drew particular attention: powered by the RK1828 AI accelerator, the dual-chip design runs dense models of up to 27B parameters smoothly, with models, files and conversation data all staying on the device, and it can connect to customers’ existing collaboration documents, private cloud drives and OA systems. It has already been deployed in several provinces. For meetings with strict data-security requirements, the offline meeting solution completes transcription, translation and minutes generation entirely on-device, so recordings and text never leave the equipment.
Progress was equally visible in robotics and home scenarios. Rockchip’s self-developed multi-scenario robot AI agent links video or image input, multimodal reasoning, tool invocation and result feedback into one closed loop, supporting autonomous scenario switching and automatic inspection-report summaries. The home-oriented RK AI NAS combines local storage with on-device computing, using a locally deployed Qwen Vision Embedding model to unify retrieval across text, photos and videos, turning a NAS box from a data warehouse into a home AI computing hub.
The value of edge AI lies in local processing of private data, zero cloud-token consumption and immediate, controllable response, which fits Rockchip’s host-SoC-plus-AI-co-processor roadmap. For the embedded and industrial-control sectors, mass-production demonstrations spanning AI office systems and robot inspection suggest that on-device LLMs are moving from proof of concept to scaled deployment, sustaining demand for hardware platforms built around RK3588/RK3576 plus RK1828.
Source: C114