Platform coverage is the question that quietly kills deployments. Everyone asks it late — after the security review passes and the budget clears — and the answer arrives as a surprise: half the field team is on Macs, the trading floor is virtual desktops, the plant runs thin clients. Ask it early, and ask it by environment.
Platforms & Environments
Does the AI App Run on Mac, Linux,
Mobile, VDI and Thin Clients?
Answered environment by environment: what is supported today, what has been demonstrated but not proven, what is out, and what containers cover on the server side.
Built from real buyer questions in our sales meetings
Direct Answer
Windows, macOS on Apple silicon, and Linux — not phones. Iternal states that AirgapAI now has full platform compatibility across all three: the macOS build is released and qualified, the Linux build runs on Ubuntu 24.04 and can run headless on a server, and Windows 10 and 11 remain the base case. Mobile is settled and negative — the products are AI PC focused only.
The limit: Windows is the most tested platform, and virtual desktops are untested at scale. Windows and macOS are built, signed and tested through separate pipelines, so macOS and Linux carry less mileage and less feature parity: Blockify data preparation sits in the Windows build, and macOS ships with standard support rather than the premium tier. On virtual desktops the only run on record is a hardware lab with no GPU, which makes that environment untested rather than unsupported.
What to verify before you sign. macOS requires an Apple M-series chip, so Intel-era Macs are out. Licenses are counted per platform, so Windows and macOS seats are bought separately. On the server side, ask for the Helm chart and its values file up front.
The client build and the server components are separate questions. The desktop application is an operating-system question. Ingestion and inference are a container question, answered with Docker, Kubernetes and a Helm chart. Most estates need both answers. For more information on where the software should run, visit the placement decision.
The Support Matrix, Environment by Environment
Buyers ask this in their own words — are you a Mac shop or a PC shop, does it also work on Linux? — and the useful answer needs three states, not two. Demonstrated sits between the other two: it has run, in a controlled environment, with no production deployment behind it.
| Environment | State | What Iternal states |
|---|---|---|
| Windows 10 / 11 | Supported | The base case and the most tested platform, down to a very old thin client. |
| macOS, Apple M-series | Supported | Released and qualified. Unified memory carries larger models than a comparable laptop. |
| Intel-era Macs | Not supported | macOS requires an Apple M-series chip. |
| Linux | Supported | Builds on Ubuntu 24.04; runs headless on a server. |
| Phones | Not supported | AI PC focused only; the products do not run on mobile devices yet. |
| Tablets | Only as an AI PC | Counts only when the tablet is itself an Intel AI PC. |
| Virtual desktops | Demonstrated | The architecture fits VDI and the Linux build enables it. The run on record is a lab with no GPU. |
| Thin clients | Supported as a pattern | Works against models and data sets hosted on a server inside your network. |
| Headless / server | Supported | A server build deploys in a Docker container with multi-user login. |
| Containers | Supported | Blockify ships as Docker containers exposing APIs, deployable in any cloud or data center. |
Read the middle column before the third. Two rows carry a hedge a datasheet would flatten into a tick.
Docker Containers and the Helm Chart
On the server side the answer is a shipped fact, not a roadmap item. Blockify, the ingestion and data-preparation layer, is delivered as Docker containers that expose APIs, hold the processing logic plus the fine-tuned models, and scale out with demand.
- One chart, values-driven. A single Helm chart deploys the platform, the same chart Iternal runs on its own Kubernetes cluster, with values files switching components on and off.
- Namespace-scoped. The chart deploys into one namespace with no globally scoped resources. You supply the container network and storage interfaces.
- Any flavor of Kubernetes. Kubernetes is the supported orchestrator and any distribution is accepted — a low bar, since on-premises teams run it already.
- Inference is a separate container. The language model is by design a service apart from the application: an NVIDIA NIM container or a vLLM container.
Ask for the chart during evaluation, not after the purchase order. The values file tells your platform team more than an architecture call will.
What Differs Build to Build
Same product, two pipelines. Windows and macOS use different signing processes, so Iternal builds and tests them separately. That is why Windows carries the deepest mileage, and why parity is worth confirming for the fleet you own.
The model list is filtered by the operating system you are on. The download page offers the macOS version or the Windows version, and the model library filters by platform and by whether the device has a dedicated GPU or only a CPU. On macOS that list includes a small 3B option, an 8B Blockify model, and a 9B build that is a real step up in capability.
Three parity details for your checklist. Licenses are counted per platform, so Windows and macOS seats are bought separately — a budget line, not a technicality. Blockify data preparation sits inside the Windows build, so a Mac-first team preparing its own data sets should plan on a Windows machine or the server-side containers. And macOS ships with standard support today rather than the premium tier.
When It Does Not Fit How Your Estate Is Built
The most common architectural objection is not about the software. It is about shape. A centralized rack-mounted unit would not work for a fully remote workforce, and large companies scaling widely use virtual desktops rather than buy laptops.
- A fully remote workforce. The objection dissolves on inspection. AirgapAI is not a rack unit; it installs on each end-user laptop and the model runs there. Distribution is the deployment model, not an obstacle to it.
- Staff and contractors reaching in. Remote users get there through VDI, remote desktop infrastructure or a VPN, and a server-side deployment still needs no internet: it runs on the local network, or over the VPN.
- A virtual-desktop-heavy estate. Here the shape genuinely changes. The Linux build is what lets AirgapAI run on virtual-desktop platforms, and the load question decides the design.
- Thin clients with nothing local to spare. AirgapAI can act as a thin client against a server holding the models and data sets, with the desktop running just the interface. One caveat: moving inference to a server does not by itself erase the requirement on the client machine.
By design AirgapAI connects only to localhost, so every server-backed pattern above turns on a light relay running on the device, pointed at the endpoint you choose.
Inside a Virtual Desktop With No Local Accelerator
Ask a virtualization architect what worries them and you hear one sentence: running the workload inside the virtual desktop makes memory and CPU consumption skyrocket. They are right to worry.
What is on record, precisely. One lab at a hardware maker has run the application with no GPU anywhere in the infrastructure, on a fraction of a Xeon CPU. That is the whole of it. Per-session figures for a production estate do not exist yet, and no published number can stand in for one — it belongs in the questions below.
Two patterns keep the session small. The conservative one moves inference out of the virtual desktop: the localhost relay sends the work to a server outside the session, leaving the desktop to carry the interface and nothing heavier. The second keeps the interface in the session while a virtual channel redirects commands to the client machine and its accelerator — scope that one with your VDI supplier first. Blockify runs on Intel Xeon CPUs with no GPU required, so an estate with no accelerators is steered, not ruled out.
Pin it down: questions for your evaluation
-
What does one session consume inside our own virtual desktop image?The only figure that matters for VDI sizing.
-
Can we see the localhost relay configured so inference runs on a server we specify?Whether the conservative VDI pattern works before you commit to a design.
-
Which AirgapAI features are in the macOS build today, and which stay Windows-only?Feature parity for a mixed fleet, in writing.
-
Send the Helm chart, its values file and the container list before we sign.Exactly what lands in your namespace, reviewed by your platform team.
Answered elsewhere
- Which devices and servers can carry the workload, and how many — see the sizing guide.
- Whether the workload belongs on the device or on your own server — see the placement decision.
- Getting the software onto a machine and a first user productive — see the setup walkthrough.
- Packaging, imaging and management tooling for a large managed estate — see the rollout page.
- How the whole deployment job fits together — see Running AI on the Infrastructure You Already Own.
FAQ
FAQ: Platforms, Devices and Delivery Environments
Yes. The official macOS build is released and qualified, and Iternal states there is now full platform compatibility across Windows, macOS and Linux. Two conditions travel with it: macOS requires an Apple M-series chip, so Intel-era Macs are out, and licenses are counted per platform.
Yes. A Linux version exists, the builds are on Ubuntu 24.04, and it runs headless on a server so it needs no user interface. That build matters beyond Linux desktops: it is what lets the application run on virtual-desktop platforms.
Not on a phone. Iternal is direct about it: the products are AI PC focused only and do not run on mobile devices yet. Tablets are the narrower case — AirgapAI runs on an Intel AI PC in laptop, desktop or tablet form, so a tablet qualifies only when it is itself an AI PC.
Not at scale. The architecture fits virtual-desktop environments and the Linux build enables the delivery, and one lab at a hardware maker has run it with no GPU in the infrastructure, on a fraction of a Xeon CPU. Treat it as untested, not unsupported.
Yes, on the server side. Blockify is delivered as Docker containers that expose APIs and run in any cloud or on-premises data center, and a single Helm chart deploys the platform, with values files switching components on and off.
Pilot the Environments You Actually Run
Build the pilot around the awkward parts of your estate, not the easy ones. Put it on a Mac and a Windows machine on day one, stand the containers up from the Helm chart with your platform team watching, and if virtual desktops carry part of your workforce, measure a session on your gold image first.