Meray AS · Sovereign Edge AI
Folkering OS runs modern AI language models locally on standard x86 hardware — no Linux, no cloud — on a small kernel written from scratch in memory-safe Rust.
Norsk utviklingslogg for Folkering OS →
Edge AI has outgrown its software. Microcontrollers cannot run modern multi-billion-parameter models. Embedded Linux can — but drags a heavy, general-purpose kernel and a copyleft (GPL) legal risk into proprietary and classified systems. Nothing was built for high-performance, secure, offline AI inference.
A bare-metal operating system written from scratch in Rust — no Linux, no libc — with memory safety and default-deny networking. The kernel contains no AI code at all: the inference engine is a separate userspace component.
It is still a research prototype. Its debug shell has no login, so it is now closed by default on real hardware and open only in lab virtual machines. Partners get the full list of known limitations.
Runs a coherent AI model locally on a modest machine — no cloud required, data never leaves the device. It ships its own from-scratch network stack (wired Ethernet and Wi-Fi) for when a mission needs connectivity — on your terms, gated by a default-deny egress firewall, not a vendor’s.
Applications and drivers run as WebAssembly on one of two paths, chosen by trust level. Mission-critical code runs fuel-metered: a frame cannot exceed its budget, and trapping is a defined, recoverable state — a property you can verify, not a promise. Large trusted workloads run JIT-compiled, where a per-frame limit would be meaningless.
A from-scratch GPU driver gives the AI workload its own hardware channel, separate from the one drawing the screen — verified against the hardware, not assumed. A kernel watchdog detects a compute kernel that never terminates, resets the engine and rebuilds the channel from cold, without human intervention. The desktop kept drawing through it.
A clean-room architecture with no Linux code lets you integrate your own proprietary or classified models — with none of the copyleft risk of Linux.
Defence · maritime autonomy · subsea · critical infrastructure. Systems that must decide locally — GPS-denied, bandwidth-denied and IP-sensitive — where the cloud is not an option.
Unless stated otherwise, these are shown in QEMU virtual machines; the laptop drivers run on the real machine.
Folkering OS and its apps are written from the ground up. The cake shows every line of code on the main branch of both repositories, counted week by week since the first commit. Drag the timeline or press Play to watch it grow. The striped slice is open-source code we borrowed (such as the Limine bootloader and the QuickJS JavaScript engine); everything else we wrote ourselves. Comments and blank lines are not counted.
Loading the numbers…
Every language model below ran on Folkering’s own inference engine, written in Rust with no machine-learning framework underneath. Speeds are measured tokens per second while the model writes its answer, on the machine named. Move the slider above to see where each run falls in the code’s growth.
Loading the model runs…
Working prototype proven on real hardware: a multi-billion-parameter model loaded and generating locally on a standard desktop machine, bit-identical to the reference run. This is a research prototype maturing towards pilot — not a finished product, and we say so on purpose. We are looking for design-partner pilots.
Request a private demo and technical briefing, under NDA: knut@meray.no.
Partners with access can read the partner wiki. How we work with partners is described (in Norwegian) on the partner page.
På norsk? En løpende utviklingslogg for Folkering OS ligger på prosjektsiden.