Moonshot AI open-sources its strongest model Kimi K3! Moore Threads domestic GPU completes Day-0 rapid adaptation

Yesterday, Moonshot AI officially open-sourced the 2.8-trillion-parameter model Kimi K3, and simultaneously released the model weights and technical report.

On the same day, Moore Threads, leveraging its AI training-inference integrated computing card MTT S5000 and MUSA software stack, took the lead in achieving rapid adaptation and stable operation of Kimi K3.

Moore Threads stated that this adaptation not only demonstrated the rapid response and full-stack collaboration capability of the MUSA software ecosystem when facing cutting-edge large models, but also once again verified the engineering strength of domestic full-function GPUs in supporting trillion-level large model innovation.

According to reports, immediately after Kimi K3 was open-sourced, Moore Threads promptly completed the model structure parsing, weight loading, core operators, cache management, and distributed runtime link adaptation.

At the same time, Moore Threads also opened up a complete adaptation path from the SGLang-MUSA inference framework to the MATE operator library, Triton MUSA, and TileLang MUSA.

In terms of underlying capabilities, the MTCC compiler and MUSA runtime can efficiently support the rapid deployment and operation of SGLang JIT and DSL operators on the MTT S5000, providing support for the stable inference of Kimi K3.

Moore Threads stated that achieving ultra-fast support for Kimi K3 marks the company's ability to accurately deconstruct cutting-edge architectures, fully connect the entire pipeline, and complete stable verification within an extremely short time.

It is understood that Kimi K3 is the world's first open-source 3-trillion-level model, and is also the most powerful Kimi model to date.

The model has 2.8 trillion parameters, natively supports visual understanding, and features a 1 million token context window, primarily targeting cutting-edge intelligent scenarios such as long-form programming, knowledge work, and complex reasoning.