Kora OS runs Gemma 4 26B at 25 tokens per second on Snapdragon NPU

Kora, a new Linux desktop operating system from a team led by Cyanogen co-founder Kirt McMaster, runs Google's Gemma 4 26B language model at 25 tokens per second for decode entirely on the Hexagon NPU inside Qualcomm's Snapdragon X2 Elite. The OS, the first product from Kora Systems Inc., previously named Playtron, is expected to arrive on PCs in 2027.

Kora splits AI workloads across the CPU, GPU and NPU, and can also distribute them across other machines on the same local network. The company says this scheduling works on the X2 Elite today through its Cortex "hybrid inference engine," a proprietary, non-open-source layer optimized for local models. The OS is built "from the kernel to the shell," runs local models by default, and calls cloud models only when needed.

McMaster told Tom's Hardware the Snapdragon X2 Elite NPU delivers "25 TPS decode throughput and 550 TPS prefill throughput" with a 32K context window. He did not specify the RAM configuration or the exact X2 Elite chip. Qualcomm pegs the Hexagon NPU at 80 TOPS.

Which NPUs does Kora actually support?

Vendor support is narrow. McMaster reiterated that "only Hexagon on Snapdragon X2 Elite is supported today." Intel NPU support is planned, with "conversations underway." AMD's XDNA 2 NPU in Ryzen AI chips is slated for launch but will "likely not run language models," McMaster said, and will instead handle embedding and speech-recognition models through ONNX.

The company's own table lays out the divide:

• Qualcomm Hexagon (Snapdragon X2 Elite): supported today via llama.cpp's Hexagon backend; runs language models.
• AMD XDNA 2 (Ryzen AI): planned via ONNX; embeddings and speech recognition only.
• Intel NPUs: planned; runtime and workloads not stated.

Each NPU vendor maintains its own software stack. Qualcomm uses QAIRT/QNN, Intel uses OpenVINO, and AMD uses Ryzen AI software. Kora bypasses Qualcomm's stack entirely. McMaster said Kora uses the "llama.cpp Hexagon backend" and "the upstream FastRPC driver" in the mainline kernel. The press kit names no Intel or AMD chip, runtime, model format, quantization or speed figure for those platforms.

llama.cpp currently offers a Hexagon backend for Snapdragon and an OpenVINO backend for Intel silicon, though the latter is marked "In Progress." There is no AMD NPU backend, which explains the ONNX route. AMD NPUs can run LLMs on Linux through the separate Lemonade 10.0 stack with FastFlowLM, as Phoronix reported in March, but Kora has not chosen that path.

How much memory will users need?

The flagship model carries real memory demands. Gemma 4 26B A4B needs roughly 14.4GB to load its weights at 4-bit Q4_0 quantization, by Google's estimate, before any context window, and third-party listings put the Q4_K_M file Kora uses at 16–17GB for download. The model activates only about 4 billion of its 26 billion parameters per token, but the full weights must sit in memory.

"We recommend 64GB because it gives you access to better models," McMaster said. That figure is a recommendation, not a firm requirement. The company projects about 70 percent of the model calls it manages will run locally in 2027.

On formats, McMaster said Kora "predominantly supports GGUF," llama.cpp's file format, and that Lemma, the inference engine inside Cortex, "can split model layers across processors as needed." Lemma is built on llama.cpp, whisper.cpp, stable-diffusion.cpp and ONNX, with more runtimes planned.

What ships first?

The desktop follows an AI-first design with "neural workspaces," sandboxed environments, and a "neural file system" that indexes files by semantic meaning. Three operating modes — Work, Create and Play — anchor the interface, with GameFrame, descended from Playtron's GameOS, covering gaming.

Humain OS, the enterprise variant developed with Saudi AI company Humain, ships first and will be "commercially available in 2027" on OEM PCs using the Snapdragon X2 Elite. The consumer Kora OS follows later in 2027. McMaster says Kora "will be posted on our site and downloadable for anyone to install," with a recommended-configuration list, and "should be installable on most modern PCs." Pricing and the business model remain undisclosed.

Kora's near-term fortunes rest on a single Qualcomm platform, with Intel and AMD support unproven until shipping runtimes and benchmarks appear.