Huawei Announces Day-Zero Ascend Support for Moonshot AI's 2.8T Kimi K3 Model
Huawei has announced immediate "0day" support for training and inference of Moonshot AI's newly open-sourced 2.8-trillion-parameter Kimi K3 model on its Ascend hardware platform. The training adaptation is achieved via MindSpeed MM on Atlas 800 and 900 A3 systems, while inference deployment is enabled through vLLM and SGLang. This rapid adaptation showcases the growing maturity of China's domestic AI hardware and software ecosystem, proving its capability to support massive 3-trillion-parameter scale models. It also highlights the industry's shift toward ultra-low-precision formats like MXFP4 to make massive model deployment more computationally efficient. The Ascend 950 supernode natively supports FP8, MXFP8, and MXFP4 numerical precision formats, enabling direct compatibility with Kimi K3's mxFP4 quantized weights. Additionally, Moonshot AI open-sourced key infrastructure libraries alongside the model, including MoonEP for expert parallelism and AgentEnv for reinforcement learning.
## BACKGROUND
Kimi K3 is a massive mixture-of-experts (MoE) model designed for advanced reasoning and long-context tasks. To train and run such large models efficiently, hardware accelerators rely on low-precision numerical formats like MXFP4 to reduce memory usage, alongside specialized communication libraries like MoonEP to balance workloads across multiple chips.