Domestic GPUs can also run trillion-parameter large models! Hygon DCU Day0 adapts to Kimi K3: 896 experts parallel without lag.

Haiguang Information announced that it has completed full-stack adaptation and performance verification of Kimi K3 on the Haiguang DCU platform, meaning domestic GPUs can now run trillion-parameter large models.

The Haiguang DCU has achieved full-stack optimization from underlying operators to inference engines, allowing the model code to run without any adjustments, with migration costs nearly zero. Developers can deploy online immediately upon obtaining computing power, demonstrating excellent compatibility.

Targeting the KDA attention and 896-expert MoE architecture, Haiguang has completed operator-level customized optimization. Even with 896 experts working in parallel, inference remains efficient and stable under heavy-load scenarios with million-level context, without lag.

In terms of scenario deployment, K3 on the Haiguang DCU can support cutting-edge scenarios such as long-range agent engineering, visual closed-loop creation, and autonomous chip design. Domestic computing power users can also run trillion-parameter open-source models.

It is worth mentioning that before releasing K3, Moonshot AI jointly completed full compatibility tuning with Haiguang and Sugon. Based on the Haiguang DCU hardware stack, Sugon has realized the distributed training and multi-card parallel inference pipeline for the Kimi MoE architecture.

The Haiguang DCU, through coordinated optimization of underlying hardware and software stack, has achieved stable operation of long-text inference on the Sugon 8000 cluster.

Official testing shows that performance loss of domestic cards is controlled within 12%, meeting the standard for commercial deployment.