en.Wedoany.com Reported - China's Rockchip showcased a "Main SoC + AI Co-processor" dual-chip collaborative solution at the 2026 World Artificial Intelligence Conference, along with a series of mass-produced terminal products, providing end-side AI industry with a complete solution from underlying chips, model adaptation, to terminal mass production.
Yang Kai, product manager of Rockchip's RK182X, stated that devices in the AIoT 2.0 era are shifting from passive recognition to active understanding and decision-making. Rockchip's dual-chip architecture consists of the RK1828 end-side AI co-processor and a main control chip (such as RK3588/RK3576), enabling deep-level optimization for the latest open-source multimodal large models ranging from 2B to 7B parameters, including Gemma4, Qwen3.5, and Qwen3.5-Omni. On-site tests showed that the first token latency for Gemma4 was as low as 169.93ms.
To reduce the difficulty of adapting domestic hardware to large models, Rockchip has developed its own RKNN3 complete development toolchain, providing one-stop support from model quantization, operator adaptation, performance tuning, to mass production deployment. In terms of ecosystem collaboration, the RK1828 has been adapted to intelligent agent solutions such as Tencent WorkBuddy and Stepfun, and is paired with the self-developed Clawchips cloud-device collaborative Agent platform.
Qiu Shi, brand manager of Rockchip's marketing department, stated that joint implementation with manufacturers like Tencent and Stepfun is a key step in improving the domestic end-side AI ecosystem. Most of the solutions exhibited this time have achieved large-scale mass production. The Hisense RGB-mini LED AI TV, built on the RK3588+RK1828, can perform real-time 2D to 3D rendering. The robotics sector showcased a cute pet companion robot, Zhiyuan humanoid robot, and the PUDU CC1 PRO commercial cleaning robot, the latter capable of cleaning up to 100 square meters per hour and now in batch deployment. The industrial edge track featured the GL-DeepSea GBox industrial quality inspection box and the Wanwu Zongheng DA600 large model all-in-one machine, the latter supporting parallel expansion with multiple RK182X modules. The RK3576M+RK1828 in-vehicle AI BOX has been equipped with the Qwen3.5-Omni full-modal model and has passed mass production verification.
Yang Kai revealed that Rockchip's next-generation end-side AI co-processor has completed core verification, achieving breakthroughs in multimodal large model computing power and local intelligent agent concurrency, and is scheduled to be unveiled in the third quarter of this year. Through modular computing power matrices, self-developed toolchains, and open ecosystem collaboration, Rockchip continues to advance the construction of an autonomous and controllable end-side AI development foundation.
