Cyanogen veteran reveals Kora, a Linux platform designed to split local AI across processors

Kora Systems Inc., led by Cyanogen co-founder Kirt McMaster, has revealed Kora, an AI-focused Linux desktop operating system designed to run models and agents locally. Built to distribute processing across CPUs, GPUs, NPUs, and other devices on a local network, the platform is scheduled to reach consumer PCs in 2027. Qualcomm Technologies spent nearly a year collaborating on the initiative, and an enterprise version developed with Saudi firm Humain, dubbed Humain OS, is set to arrive commercially on Snapdragon X2 Elite OEM systems earlier in 2027. [1][2][3]

At present, hardware support is limited. McMaster told Tom's Hardware that only the Hexagon NPU on Qualcomm's Snapdragon X2 Elite is currently supported, delivering 25 tokens per second decode and 550 tokens per second prefill throughput on Google's Gemma 4 26B model with a 32K context. The system relies on a proprietary engine named Cortex, which leverages Lemma—a stack built with llama.cpp, whisper.cpp, stable-diffusion.cpp, and ONNX—to split GGUF model layers. While support for Intel and AMD hardware is planned, McMaster indicated AMD NPUs will likely be restricted to speech and embedding workloads rather than language models. [1]

Kora projects that roughly 70 percent of managed model calls will run locally by 2027, defaulting to cloud services only when necessary. [1]