
Kora Linux AI OS: Local LLMs Across CPU, GPU and NPU
Kora is a Linux desktop OS that runs local AI by default and splits models across CPU, GPU and NPU. Gemma 4 26B hits 25 tokens per second on X2 Elite.
Most AI on the desktop today lives inside an app or a browser tab. Kora takes the opposite approach: it puts the models inside the operating system itself. Unveiled on November 10 by Kora Systems Inc., the company formerly known as Playtron and led by Cyanogen co-founder Kirt McMaster, Kora is a Linux-based desktop OS that runs local AI models by default and schedules them across the CPU, GPU and NPU, and even across other machines on your local network.
- What it is: A Linux desktop OS with built-in models and agents, organized around sandboxed "neural workspaces"
- Local AI claim: According to McMaster, the Snapdragon X2 Elite's Hexagon NPU runs Gemma 4 26B at 25 tokens per second decode and 550 tokens per second prefill with 32K context
- Runtime: Its Lemma inference engine builds on llama.cpp, whisper.cpp, stable-diffusion.cpp and ONNX, and mainly uses GGUF models
- Timing: The enterprise variant ships on OEM PCs in 2027, with Kora itself following later in 2027 as a free download
How Does Kora Run Local LLMs?
The core is Cortex, which Kora calls a hybrid inference engine. It runs local models first and only reaches for cloud models when a task needs them, and the company projects that about 70 percent of the model calls it manages will run locally in 2027. Inside Cortex, the Lemma engine can split model layers across processors as needed, McMaster told Tom's Hardware.
On Qualcomm hardware, Kora reaches the NPU through llama.cpp's Hexagon backend and the upstream FastRPC kernel driver rather than Qualcomm's own QAIRT runtime. That is a notable choice for anyone who follows the local LLM hardware world, because it builds on the same open-source tooling many self-hosters already use. It's FOSS reports that Qualcomm engineers worked alongside the Kora team for nearly a year to bring the Linux desktop to the Snapdragon X2 Elite.
Which NPUs Does Kora Support?
Today, only the Hexagon NPU on Snapdragon X2 Elite is supported, which Qualcomm rates at 80 TOPS. AMD XDNA 2 support is planned for launch through ONNX, and Intel NPU support is in discussion. McMaster said neither Intel's nor AMD's NPU will be used for language models, since GPUs handle LLMs better on x86 machines. Instead, those NPUs will likely run embedding and speech-recognition models. That split is practical engineering: give each accelerator the jobs it does best.
How Much RAM Does Kora Need?
Memory is the real constraint for local AI on any compact PC. Tom's Hardware notes that Gemma 4 26B A4B needs about 14.4GB just to load its weights at 4-bit quantization by Google's estimate, and the Q4_K_M file Kora uses is around 16 to 17GB to download. Because it is a mixture-of-experts model, only about 4 billion of its 26 billion parameters are active per token, which is why the throughput is so usable. Kora recommends 64GB of RAM for access to better models, though it says that is not a firm requirement. If you are weighing machines for this kind of workload, our Mac Studio vs DGX Spark comparison breaks down why unified memory matters.
Workspaces, Privacy and Play
Beyond inference, the desktop is built for keeping AI context contained. Each neural workspace keeps a project apart, the AI only sees the open workspace, files keep their own version history, and AppArmor enforces which files, devices and services each app and agent can touch. A semantic "neural file system" lets you search by meaning. It's FOSS notes the memory system analyzes connected email, calendar and files with local models, sending data to the cloud only with user approval. Gamers are covered too: Playtron's GameOS lives on as GameFrame, a one-click console mode.
One caveat for open-source fans: Cortex itself is not open source, though the developer building blocks around it are. With no ISO available yet, Kora is a 2027 story, but it is one of the most interesting bets yet on local-first AI for compact Arm PCs. More hardware news lives in our mini computers section.
Sources: Tom's Hardware — November 11, 2026; It's FOSS — November 10, 2026.
More Mini Computers Stories

ESP32-P4 Desktop OS: Sylphide OS and BMDS Turn MCUs Into PCs
Two open-source desktop OS projects, Sylphide OS and BMDS, give the $25 Olimex ESP32-P4-PC windows, apps, SSH and even Doom at 1280x720 and 60 FPS.

Tektronix 1 Series Oscilloscope: $450 Scope With AI Control
Tektronix's 1 Series DPO starts at $450 with 12-bit resolution, up to 200 MHz, I2C/SPI/UART decoding and an AI connector for natural-language control.

Eight Used V100 GPUs: How a Lawyer Built a Private AI Server
A lawyer's private AI server runs DeepSeek V4.1 Flash on eight used V100 GPUs at up to 215 tokens per second, keeping case files off the cloud.
