Kora Systems, the Palo Alto company formerly called Playtron and led by Cyanogen co-founder Kirt McMaster, unveiled Kora on Nov. 10, a Linux desktop OS whose Cortex inference engine splits AI workloads across the CPU, GPU and NPU and across other machines on the same network.
Cortex runs local models by default and adds frontier cloud models when a task calls for them, according to Kora’s press kit, which projects that about 70% of the model calls it manages will run locally in 2027. The engine is proprietary and not open source. The OS underneath is built on Linux and Nix.
NPU support is limited to one chip. The 80 TOPS Hexagon NPU in Qualcomm’s Snapdragon X2 Elite is the only one Kora uses, McMaster told Tom’s Hardware.
“only Hexagon on Snapdragon X2 Elite is supported today.”
Kirt McMaster, CEO, Kora Systems
Intel and AMD NPUs will not run language models; on x64 machines, those go to the GPU. Intel NPU support is planned, and on AMD Ryzen AI chips the NPU will likely handle embedding and speech-recognition models through ONNX. On the X2 Elite, McMaster said, Gemma 4 26B decodes at 25 tokens per second entirely on the NPU with a 32K context. The company recommends 64GB of RAM but does not require it.
Humain OS, an enterprise edition built with Saudi AI company Humain and first shown at LEAP 2026 in Riyadh, ships first, with commercial availability set for 2027 on OEM PCs using the X2 Elite. Kora itself follows later in 2027, according to the Kora and Qualcomm joint release, which also says Qualcomm is upstreaming X2 Elite support into the mainline Linux kernel.
The Humain and Kora release lists Samsung, Square Enix, Polychain, VanEck, Circle and Alumni Ventures as backers, all of them carried over from Playtron; the announcement included no new funding round. About 50 engineers are building the OS with AI coding tools, including OpenAI’s Codex and Anthropic’s Claude, TechCrunch reported, a working pattern that matches current vibe coding adoption figures.
Llama.cpp as the NPU path
Kora reaches the Hexagon NPU through llama.cpp’s Hexagon backend and the upstream FastRPC kernel driver, not through QAIRT/QNN, Qualcomm’s own runtime. McMaster said Kora mainly supports GGUF models. Lemma, the engine inside Cortex, is built on llama.cpp, whisper.cpp, stable-diffusion.cpp and ONNX, so the closed layer sits on open runtimes and a mainline driver.
