Kora Linux OS distributes AI workloads across chips and local machines

A team led by Cyanogen co-founder Kirt McMaster has unveiled Kora, a Linux desktop operating system designed to run AI models locally by default, using cloud models when needed. Its proprietary Cortex inference engine can distribute workloads across CPUs, GPUs, NPUs and other computers on the same local network. The desktop includes sandboxed workspaces, semantic file indexing and gaming capabilities inherited from Playtron’s GameOS.

Only the Qualcomm Hexagon NPU in the Snapdragon X2 Elite is currently supported. McMaster says it runs Google’s Gemma 4 26B model at 25 tokens per second for decoding, although the system’s RAM configuration and specific chip were not disclosed. Intel and AMD NPU support is planned, but neither will handle language models in Kora because the developer considers GPUs better suited to those workloads on x64. AMD NPUs will likely handle embedding and speech-recognition models through ONNX. Cortex’s underlying Lemma engine can split model layers across processors and builds on several open-source inference projects.

The company recommends 64GB of memory for access to better models, without making it a requirement. Its enterprise variant, Humain OS, developed with Saudi AI company Humain, is scheduled to arrive first on Snapdragon X2 Elite PCs in 2027. Kora will follow later in 2027 as a downloadable installation for most modern PCs. Pricing and the business model remain undisclosed.