Moore Threads MTT S5000 GPU Runs Kimi K3 2.8T Model on Day 0

Moonshot AI released Kimi K3, a 2.8 trillion parameter model, and Moore Threads matched that launch with immediate GPU support. This pairing matters because running such a massive model requires specialized hardware and software stacks that few vendors can handle on day one. Users and enterprises relying on domestic Chinese silicon now have a verified path to deploy this specific large language model without waiting for third-party drivers.

Moore Threads MTT S5000 GPU supports Kimi K3 2.8 trillion parameter model

The hardware anchor for this release is the MTT S5000, a full-function AI training and inference card from Moore Threads. Moore Threads built its adaptation around the MUSA software stack, which serves as the bridge between the Chinese GPU architecture and the AI model. This setup allows the MTT S5000 to execute the complex computational demands of the Kimi K3 architecture directly.

Kimi K3 operates with a 1 million token context window and includes native visual understanding capabilities. Moore Threads supports this workload using the SGLang-MUSA inference framework. The software layer also integrates the MATE operator library, Triton MUSA, and TileLang MUSA to manage the model's heavy computational load efficiently.

Moore Threads completed the adaptation on the same day Moonshot AI open-sourced the model weights and technical report. The engineering team handled model structure parsing, weight loading, core operator mapping, cache management, and distributed runtime chain adaptation. The company states that this rapid integration validates the engineering strength of domestic full-function GPUs in supporting trillion-parameter models.

We looked at the last Kimi K3 update while tracking these developments, noting that stability remains a key theme for this model line. Moore Threads has confirmed that the MTT S5000 runs Kimi K3 stably with no reported crashes during initial deployment. This provides a concrete, working reference for organizations evaluating domestic GPU alternatives for large-scale AI inference tasks.