AI Model Report

Infrastructure · NOVEMBER 13, 2026

KORA puts local AI inside Linux—but its desktop remains a 2027 plan

Cortex promises inference across CPU, GPU and NPU. For small businesses considering sensitive customer workflows, local-by-default is not the same as cloud-free.

By Aiko Tanaka · Inference & serving · November 13, 2026

KORA Systems unveiled its Linux-based desktop operating system on November 10, saying its Qualcomm collaboration enables a 26-billion-parameter language model to run entirely on Snapdragon X2 Elite PCs. KORA is scheduled for 2027, not available for deployment today.

For small-business owners considering sensitive customer intake or support, the distinction matters: local inference could reduce reliance on cloud models. Local-by-default doesn’t establish that a workflow is cloud-free.

What Cortex schedules

The KORA announcement and Qualcomm joint release describe Cortex as an inference engine scheduling workloads across CPU, GPU and NPU, plus other machines on the local network. They also identify on-device speech-recognition and embedding models.

Tom’s Hardware obtained implementation details from CEO Kirt McMaster. Cortex’s internal inference engine is Lemma; the reported model stack includes llama.cpp, whisper.cpp, stable-diffusion.cpp, ONNX and the GGUF model format. McMaster said KORA uses llama.cpp’s Hexagon backend and the upstream FastRPC driver.

The Snapdragon X2 Elite has an 80 TOPS Hexagon NPU. At the time of reporting, only Hexagon on X2 Elite was supported, a narrower implementation than the broader scheduling description. Support for Intel or AMD acceleration wasn’t established.

What the performance claim leaves out

McMaster told Tom’s Hardware that Gemma 4 26B achieves 25 tokens per second in decode and 550 in prefill, entirely on the Hexagon NPU with a 32K context. These are vendor-reported figures, not independent benchmarks.

Both matter to an operator assessing an interactive customer-facing workflow, but the figures don’t establish its accuracy or suitability.

Tom’s Hardware noted that neither the RAM configuration nor the specific X2 Elite chip was disclosed. McMaster recommends 64GB of RAM for access to better models, while saying that isn’t a firm requirement.

WIRED reports support for x86-64 PCs as well as Snapdragon. The joint release claims selected Windows x86-64 applications will work on Snapdragon, not the entire Windows catalog. Hardware support and application compatibility remain separate checks for prospective users.

A 2027 desktop, not a deployment decision

TechCrunch reports that KORA uses Nix, replacing GameOS’s Fedora Silverblue foundation. Cortex and KORA’s desktop interface remain proprietary.

WIRED describes semantic file search that retrieves documents by contents rather than filenames. Cofounder Franck De Girolami told WIRED:

"All your context, all your memory is on your machine,"

That statement frames the local-workspace promise. Descriptions of memory, permissions and SDK features don’t establish whether a workflow using cloud models transmits business data. For an owner handling prospect or customer information, that remains a separate question from where a model runs.

The joint release places HUMAIN OS first on Snapdragon PCs in 2027, with KORA following later that year. Phoronix likewise distinguishes the announcement from a release.

For a small team, locally processed intake or support is a potential use, not a demonstrated customer-acquisition result. A deployment decision should wait for availability, independent performance testing, application checks and evidence about data handling.

Sources