On November 10, 2026, KORA Systems announced KORA, a new Linux-based desktop OS. It is aimed mainly at developers and creators, and a release for PCs is planned for 2027. Its most distinctive feature is that the OS itself includes "Cortex," an engine that manages AI processing, along with a memory function, so that AI models and agents can share context about a user's work. According to the company, users can switch between multiple AI models while continuing the same work.
At the same time, AI access to information is restricted by permissions, and Cortex itself is not open source. A wide choice of AI models, control over what information is handed to AI, and the ability to verify the internals of the runtime from the outside are three separate things and need to be considered individually.
Cortex Runs a 26-Billion-Parameter AI Model Locally
According to KORA Systems' announcement, Cortex distributes AI workloads across the CPU, GPU, and NPU, and can also draw on the computing resources of other machines on the same local network. By default it uses AI models running on the device, and it is designed to use cloud-based models when needed. Cortex is not itself an AI model. Its role is to manage multiple models and computing resources and route each task to the appropriate place to run.
The company has spent about a year co-developing a Linux desktop with Qualcomm built around the Snapdragon X2 Elite. According to the two companies, this platform, which combines an Oryon CPU, an Adreno GPU, and a Hexagon NPU rated at 80 TOPS, can run a 26-billion-parameter language model locally. They say it also supports speech recognition and embedding models.
However, 80 TOPS is a measure of compute performance, while 26 billion is the number of parameters in an AI model, so the two figures cannot be compared directly. The announcement also did not disclose the name of the model that was run, its quantization method, or the memory capacity required. Generation speed and power consumption have not been published either, so how well it performs in real-world use will have to wait for independent testing.
Managing AI Memory and Access Permissions by Workspace
KORA's "Neural Workspace" separates work environments, such as work versus personal use, or by client or project. Permissions for apps and AI agents are set for each workspace, and the system also uses AppArmor, a Linux access control mechanism. The design is meant to control access to files, hardware, AI tools, external services, and more.
According to KORA Systems, local AI models can analyze contacts, email, calendars, cloud-stored files, and other data that the user has authorized, and keep the work history and context on the device. This information is designed to carry over between multiple AI agents. For example, the company says that even if a user switches from the same chat to Codex, Claude Code, Gamma, or other tools, the context of their work is preserved within the permitted scope. The aim is to reduce the effort of re-explaining past work and background every time the user switches AI tools.
Supporting this is the company's "Neural File System." In addition to conventional file names and storage locations, it builds an index based on the meaning of files and manages change history at the OS level. The design also lets developers use project information, memory, semantic search, and other features through the SDK, following each workspace's permission settings. How these features will be available in individual apps, however, will have to wait for further details.
Storing work memory on the device does not mean that no information is ever sent outside. In its Q&A, KORA Systems explains that relevant context is sent to cloud services only when the user explicitly permits it. In other words, local information management is the baseline, but information can be shared with external AI services if the user allows it.
AI Models Are Freely Selectable, but Cortex Is Closed
KORA is designed not to depend on any particular AI model. However, being able to choose models freely, being able to control what information each model can access, and having the source code of the AI runtime engine be public are separate matters. The company states clearly that Cortex is KORA's proprietary technology and is not open source.
The design announced so far can be summarized as follows.
| Item | KORA's design | Current limitations and caveats |
|---|---|---|
| AI model selection | Policy of allowing users to choose local and cloud AI models regardless of developer | Support for and operation of every AI model is not guaranteed |
| Access to information | Controlled through workspaces and individual permission settings | Permitted context can also be sent to the cloud, so processing is not necessarily fully local |
| Transparency of the AI runtime engine | Cortex is developed as KORA's proprietary technology | Source code is not public, so the internal implementation cannot be directly verified from outside |
What matters to users is whether the AI models they already use are supported, which information they allow each model to access, and how far they can check the OS's permission management. The fact that Cortex is not open source cannot by itself be used to judge its safety, but there are limits to the transparency of its internal implementation. Likewise, the separation between workspaces and the permission controls are, for now, an announced design, and how well they work in an actual product has not been verified.
HUMAIN OS to Launch First in 2027, With KORA Following the Same Year
KORA's technology base is expected to be adopted first in an enterprise product. "HUMAIN OS," which KORA Systems is co-developing with HUMAIN, shares the same Linux foundation and AI agent technologies as KORA, and is planned for commercial launch on OEM PCs in 2027. According to a joint announcement by KORA Systems and Qualcomm Technologies, KORA will become available later in 2027, after HUMAIN OS ships.
For existing Windows apps, the company says that selected x86-64 apps will run on Snapdragon-based PCs through compatibility and translation technology. This does not guarantee support for all Windows apps. The announcement also did not include a list of supported apps, and specific OEM manufacturers, models, and pricing have not been disclosed.
KORA is built on GameOS, which the company developed. It also plans to include a feature for switching to a game-oriented screen through "GameFrame," which comes from GameOS. However, for it to become widely used as an everyday OS for developers and creators, compatibility with existing apps, development tools, and AI models will be essential, not just gaming-related features.
KORA is aiming for a new PC environment in which the OS itself manages the context and permissions of work, rather than treating AI as just another application. Ahead of the 2027 product launch, what will matter, beyond supported apps and pricing, is how safely and smoothly work memory can be carried across different AI models. Only once the practicality of these features is confirmed can its value as an alternative to existing Windows and Linux environments be properly assessed.