Gasgoo Munich- On November 5, ACE ROBOTICS has partnered with HYGON to complete the adaptation, tuning, and deployment verification of its Kairos 3.1 world model on HYGON's Deep Computing No. 3 DCU accelerators and C86 CPUs. The move clears the path for deploying the model on domestic computing platforms.
Image source: ACE ROBOTICS
In 480P text-to-video inference tasks, a single HYGON Deep Computing No. 3 DCU card took 12.42 seconds to process data. That performance reaches 94.2% of comparable mainstream international GPUs — a 32.2 percentage point increase from the 62% achieved during the initial adaptation, while maintaining essentially identical video generation precision.
Leveraging the HYGON DTK software stack, the partners optimized the software dependencies and core operators required for the model, completing two rounds of compatibility certification. The solution currently supports both all-in-one appliances and cluster deployments, allowing for implementation via pre-installed model images.
Unveiled at the World Artificial Intelligence Conference in July, Kairos 3.1 uses a hybrid Transformer architecture that fuses understanding, generation, and prediction capabilities. It is designed for tasks such as robotic manipulation and physical scene simulation.
On the application front, ACE ROBOTICS has gathered data covering tens of thousands of SKUs across five stages and 13 atomic actions in instant retail warehousing, with model-generated scenes already feeding into actual training. In home service scenarios, Kairos has achieved edge deployment, supporting long-horizon action interactions such as retrieving items from refrigerators and passing objects.
While this adaptation opens a new technical path for scaling world models on domestic computing platforms, performance under varying workloads and real-world application environments still requires further verification.
