Kora's Linux Desktop Splits AI Across CPU, GPU and NPU, but Only Snapdragon X2 Elite Works Today

Kora is an ambitious pitch: a Linux desktop built from the kernel up with AI models and agents that can spread work across a PC's CPU, GPU and NPU, and even across other machines on the same local network. The company behind it, Kora Systems Inc., is led by Cyanogen co-founder Kirt McMaster, and the product grew out of Playtron. The headline capability is real, but the fine print matters a lot. Right now, one chip is supported, and the cross-vendor story is still a plan.

What Kora is trying to be

Kora calls itself "The Operating System for Intelligence," and the marketing leans into that idea throughout. The desktop uses "neural workspaces," which are sandboxed areas for tasks, and a "neural file system" that indexes files by semantic meaning rather than just by name or folder. Three operating modes, Work, Create and Play, organize the experience. Play is handled by GameFrame, which descends from Playtron's GameOS.

The engine underneath is Cortex, which the company describes as a hybrid inference engine. Cortex isn't open source, and it's tuned for local models. By default, Kora runs models on the device and only reaches for cloud models when a task needs them. Humain OS, developed with Saudi AI company Humain, is the enterprise variant of the same stack. The company expects Kora to reach PCs in 2027.

Hexagon is the only NPU on the list

The release material names one NPU: the Hexagon NPU in Qualcomm's Snapdragon X2 Elite, which Qualcomm rates at 80 TOPS. On that chip, Cortex can schedule AI workloads across all three processors and across other machines on the network. Cross-machine scheduling is the more unusual idea here, and it's the part that could matter most if it works beyond Qualcomm hardware. For now, McMaster has been direct: "only Hexagon on Snapdragon X2 Elite is supported today."

Intel NPU support is planned, and McMaster said "conversations are underway." He also clarified that neither Intel's NPU nor AMD's will handle language models. AMD's Ryzen AI chips are expected to use the NPU for embedding and speech-recognition models through ONNX. Neither Intel nor AMD chips are identified with a specific NPU runtime, model format, quantization level or speed figures in the material.

Why each NPU needs its own runtime

The vendor split is the main obstacle to a single cross-vendor AI layer. Each NPU vendor has its own software stack and prefers its own compiled model format: Qualcomm uses QAIRT/QNN, Intel uses OpenVINO, and AMD uses Ryzen AI software. Windows testing of AI PCs has shown the same divide. Snapdragon reached its NPU through the QNN provider under ONNX Runtime, while Intel's workloads ran on the GPU through IPEX because the GPU beat the NPU on speed.

Kora's route on Snapdragon goes around Qualcomm's own runtime. McMaster said the system uses the "llama.cpp Hexagon backend" and "the upstream FastRPC driver," which points to the open-source llama.cpp project reaching the NPU through the FastRPC driver in the mainline kernel. That's a notable choice. It means Kora's Hexagon support depends on community-maintained code, not Qualcomm's stack, and that code will need to keep pace with new silicon.

Model formats and the Lemma layer

On model formats, McMaster said Kora "predominantly supports GGUF," the llama.cpp file format. The inference engine inside Cortex, called Lemma, "can split model layers across processors as needed." Lemma is built on llama.cpp, whisper.cpp, stable-diffusion.cpp and ONNX, and the runtime chosen depends on what each model needs. More runtimes are promised later.

The upstream picture helps explain the choices. llama.cpp has a Hexagon backend for Snapdragon and an OpenVINO backend that targets Intel CPUs, GPUs and NPUs, though the OpenVINO backend is still marked "In Progress." There's no AMD NPU backend, which likely explains why AMD goes through ONNX. AMD NPUs can run LLMs on Linux through a separate route, the Lemonade 10.0 stack with FastFlowLM, which Phoronix reported in March. Kora hasn't taken that path.

The numbers and the memory question

McMaster said the X2 Elite reaches "25 TPS decode throughput and 550 TPS prefill throughput" with the Gemma 4 26B model, running "entirely on the Hexagon NPU with 32K context." The figures are the company's own, and the material doesn't specify the RAM configuration or the exact X2 Elite chip, so they're hard to compare directly with other setups.

The model is a mixture-of-experts design. Google lists it as Gemma 4 26B A4B, and it activates roughly 4 billion of its 26 billion parameters per token. That keeps per-token compute low, but every weight still has to sit in memory. Google estimates about 14.4GB to load the weights at 4-bit (Q4_0) quantization, before any context window. Third-party listings put the Q4_K_M file that Kora uses at roughly 16 to 17GB. McMaster noted that a dense model of the same size would run far slower on this hardware, which is the reason the MoE design matters here.

On hardware, McMaster said "we recommend 64GB because it gives you access to better models," but he didn't present that as a requirement. He also said Kora should be "installable on most modern PCs," with a list of recommended configurations to come. Pricing and the business model will be shared closer to launch.

What to watch

The company projects that "about 70 percent of the model calls it manages will run locally in 2027." That's a bold target, and it depends heavily on how many NPUs can run useful models. Right now, the answer is one chip family, with Intel and AMD still in planning. Humain OS is slated to arrive first on OEM PCs with the Snapdragon X2 Elite and is described as commercially available in 2027, with the consumer Kora desktop following later that year.

The cross-machine scheduling is the item to track. If Cortex can split a model across a laptop, a desktop and a second box on the same network, that's a meaningful capability. It's also one that depends on the runtime layer working across mixed hardware, which is exactly where Kora's current support is thinnest. Until Intel and AMD backends ship and get benchmarked on the same models, the real question is whether Kora's one-chip start can grow into the multi-vendor platform the pitch describes.