Gasgoo Munich-GigaAI and Moore Threads have struck a strategic partnership at this year's World Artificial Intelligence Conference (WAIC) 2026. The companies announced that GigaAI's embodied world model has successfully completed training and inference verification on Moore Threads' full-function GPUs, delivering results that align closely with mainstream computing platforms.
Image Source: GigaAI
This partnership signals that the computationally intensive "world model training" phase of the physical AGI era can now operate stably on domestic computing infrastructure. It validates a fully autonomous, full-stack pathway combining domestic models with domestic chips.
The partnership rests on the computing power of Moore Threads' flagship MTT S5000, a card designed for both training and inference. Built on the proprietary MUSA unified system architecture, the S5000 integrates diverse capabilities including AI acceleration, 3D graphics rendering, and physics simulation. It supports full-precision computing from FP8 to FP64 and is the first of its kind to pass national-level security and reliability assessments.
During the verification process, GigaAI's core model architecture—encompassing its embodied base model and driving world model—demonstrated full capability for both training and inference on the MTT S5000. More importantly, the results aligned closely with international mainstream solutions, showing strong efficiency per unit of computing power.
As the first Chinese tech company to systematically architect world models, GigaAI has established a four-pillar, full-stack strategy: "embodied base model—world model—native ontology—generalized scenarios." Its embodied base model, GigaBrain-0, took the global top spot in RoboChallenge, outperforming Physical Intelligence's π0.5 by nearly 10 percentage points. Meanwhile, its world model platform, GigaWorld-Policy, claimed the championship in the authoritative RoboCasa365 evaluation, defeating rivals like Nvidia's GR00T N1.5.
Looking ahead, the two companies plan to deepen their strategic partnership using this verification as a foundation. Their goal is to build the entire "training—inference—deployment" pipeline on domestic software and hardware. Together, they aim to create a global benchmark for autonomous, controllable general world models—laying the groundwork for the large-scale deployment of physical AGI on domestic computing infrastructure.
