Moonshot AI open-sources Kimi K3 with 2.8 trillion parameters, Huawei Ascend adapts on day one

Moonshot AI officially open-sourced its latest Kimi K3 model on July 27, with 2.8 trillion parameters, making it the first open-source model at the 3 trillion scale globally. The model is designed for advanced AI tasks like long-form coding, knowledge work, and reasoning.

Huawei announced that its Ascend platform achieved zero-day adaptation right after the release. On the training side, it completed basic adaptation on Atlas 800 A3 and Atlas 900 A3 SuperPoD using MindSpeed MM, with optimizations for operators, parallel processing, and memory. For inference, it supports quick deployment via vLLM Ascend and SGLang engines. The Ascend 950 supernode supports full numeric precision formats including FP8, MXFP8, and MXFP4, natively handling Kimi K3's mxFP4 quantized weights.

Kimi also revealed three Infra technologies behind the model: MoonEP, FlashKDA, and AgentEnv, covering key areas from high-performance communication and operators to distributed RL environments. FlashKDA was already open-sourced, while MoonEP and AgentEnv are now released with this model.