On July 27, 2026, Moonshot AI officially open-sourced Kimi K3. The model has a total of 2.8 trillion parameters, making it the world’s first open-source model in the 3-trillion-parameter class. It is designed for advanced intelligent scenarios including long-horizon programming, knowledge work, and reasoning.
Huawei officially announced today:
Immediately after Kimi K3 was open-sourced, Ascend achieved rapid day-0 adaptation. On the training side, foundational support for layer-reduced training was completed on the Atlas 800 A3 and Atlas 900 A3 SuperPoD based on MindSpeed MM, with dedicated optimizations for operators, parallel acceleration, and memory usage. On the inference side, rapid deployment is enabled through the open-source inference engines vLLM Ascend and SGLang. Meanwhile, the Ascend 950 SuperPoD supports the full range of numerical precision formats, including FP8, MXFP8, and MXFP4, and natively supports Kimi K3’s mxFP4 quantized weights.
According to an IT Home report yesterday, Kimi officials also announced three Infra technologies supporting Kimi K3 training: MoonEP, FlashKDA, and AgentEnv, covering key links ranging from high-performance communication and high-performance operators to distributed RL environments. FlashKDA had already been open-sourced, while MoonEP and AgentEnv were officially open-sourced with this release.
Open-source links:
Hugging Face: https://huggingface.co/moonshotai/Kimi-K3
GitHub: https://github.com/MoonshotAI/Kimi-K3