NightRun: an LLM that boots from a USB stick with no operating system, written almost entirely by a coding agent

By Zendoric · September 6, 2026.

An open source project called NightRun has pulled off something unusual: a language model that boots directly from a USB stick or an SD card, with no operating system in between. No Linux, no Windows, not even a generic kernel running underneath. According to its own documentation, the machine's firmware (UEFI, the software that replaced the old BIOS and prepares the hardware before handing control to an operating system) boots NightRun directly, as if it were the system itself.

The program — written in Rust, a language designed to prevent memory errors in low-level code — loads the entire model into RAM (between 1.3 and 2.4 GB depending on the model chosen), verifies its integrity with checksums as it reads, and then seals the storage: it never touches it again. The conversation takes place over a graphical terminal drawn by the runtime itself, without help from any external operating system's compositor. It works on x86_64 PCs via USB and on the Raspberry Pi 5 via SD card, where, with no operating system present, the program itself has to handle even fan control.

There is no network stack: that is a design decision, not a temporary limitation. Once the model is loaded, the machine can neither receive nor send anything, as the project itself describes. NightRun currently supports four small, quantized models — that is, with their parameters compressed to lower numerical precision so they take up less memory and run faster — of between 1 billion and 4 billion parameters: two versions of Llama 3.2 (Meta), Granite 4.1 3B (IBM) and Qwen3 4B (Alibaba). The repository's own benchmarks, measured rather than projected as the project clarifies, put performance in a modest range: between 3 and 20 tokens per second in generation depending on the model and the hardware, with markedly lower figures on a real Raspberry Pi 5 than in a virtualized environment with eight cores.

The most striking detail, however, is not about performance but about authorship. As the project itself acknowledges, most of the code was written with Claude Code, Anthropic's coding agent, using the model NightRun calls "Fable 5." It is no minor detail: the author presents it as a deliberate test of how far a coding agent can be pushed when the target is a systems project with "hard correctness and safety requirements," code that boots before any operating system, with no safety net if something fails.

That is precisely the terrain where coding agents used to fail most: low-level programming, with no automatic memory manager, with compute kernels hand-written for vector instructions (AVX2 on x86 processors, NEON on ARM) and no room for subtle errors. The methodology the project describes — validating every output token by token against llama.cpp (the community's reference engine for running these models), adversarial testing of the file format parser, and an installer that demands exact, written confirmation before erasing a disk — is what you would expect of any human systems engineer. That an agent can sustain it on a project of this kind is, more than the novelty of NightRun itself, the underlying story.

It is worth keeping this in proportion. As far as its own website indicates, NightRun is a personal open source project, with no company or funding round behind it, and minimal traction so far: it reached Hacker News with a single vote and no comments. Its performance figures — a few tokens per second with models of 1 billion to 4 billion parameters — do not compete with anything running in the cloud or with the frontier models we routinely evaluate at Zendoric. It is, above all, an engineering curiosity, not a product with market ambitions.

But that curiosity connects with two threads that do matter to us. The first is local AI sovereignty: the current pushing artificial intelligence onto the user's own hardware, with no intermediaries and no cloud, which at Zendoric we have flagged as one of the healthiest democratizing forces in the sector. NightRun takes that idea to its logical extreme — a device that, by construction, cannot leak data because it has nothing to leak it with — a design with real applications in environments that demand total isolation: critical systems, air-gapped devices, embedded electronics where the guarantee that "this connects to nothing" is worth more than the convenience of a full operating system.

The second thread is the capability of coding agents themselves. We have been tracking their progress on benchmarks such as SWE-bench and Terminal-Bench, which assess conventional software development tasks. NightRun is an isolated data point, not a controlled test — there is no way to independently verify how much of the project is the agent's work versus human supervision — but it points in a direction we already suspected: coding agents are beginning to reach into domains that previously required very specific specialists, such as firmware or low-level kernels. If that trend is confirmed with larger and more thoroughly audited projects, the medium-term effect is not that systems engineers disappear, but that their work shifts: fewer hours writing repetitive code and more time supervising, verifying and deciding what is worth building. It is the same reading we hold for the rest of the labor market exposed to AI — work changes in nature before it shrinks in volume — applied, this time, to those who write the software that makes everything else possible.