AienOS

Persistent machine intelligence on hardware you own

What AIEN is · The machines · Speed, with receipts · Right now · The tree · Research · How a change gets in · Principles · Licence

Aien is a computing stack built to run on machines you own, with nothing in the trusted base that depends on an outside service: its own operating system kernel (Aienux), its own language and GPU engine (Omega), a host runtime and model server in Rust, and a verification pipeline that retains evidence for qualified runs. AienOS is now in alpha testing (since 2026-10-26): testers install the alpha bundle (the Core engine, the Omega install wizard and the CUDA++ user-space replacement), whose GPU path is qualified only on the DGX Spark GB10; components outside that bundle keep their own pre-alpha labels. It started on one NVIDIA DGX Spark and now runs across several machines: the Spark serves a 27B model on our own GPU engine, an Intel iMac from 2017 boots the Aienux kernel on bare metal, a Raspberry Pi 5 has run the same kernel on real hardware and hosts the project's house server, QEMU carries the emulator gates, and a Windows PC with an RTX 5070 is the target for AIEN Rig. The project is led by one engineer, Drake Stapleton, working with AI agents as day-to-day collaborators: the agents write, test and review; the engineer decides. AienOS is the name for the core tree these parts are being brought into. Every claim on this page links to the commit, pull request or evidence folder it comes from.

A number here stands beside two things only: the machine's physical ceiling and our own previous run. We do not compare with anyone else's system. Every 27B number below is on a branch today; nothing is on main until the alpha merges land.

Numbers still being worked are not in this table: continuous batching on the 27B door (about 7.0 tokens/s single door, batch runs pending, receipt) and the MTP chain (5.80 tokens accepted per pass at k=8, wall speed not yet reported, receipt).

The live ledger is at www.aienos.com/now/. It renders the project's coordination ledger as it is written: every line is a promise with an owner and an observable finish, agents write their progress there, the operator writes decisions there, and nothing on it is composed by hand. The streams in motion today:

• Engine speed cascade on the 27B: a tree-verify GEMM, an MTP chain with a second drafter, 4-bit decode, continuous batching, prefill kernels, document memory up to 32k, and a drafter model; each lands as a receipt on a branch of omega-mv and aien-sovereign-core-mv.
• Alpha testing: the alpha bundle (Core engine built from aien-sovereign-core-mv main, the seeded Omega install wizard and omc, the CUDA++ user-space replacement) ships through GitHub Releases on aien-dev/aienos-releases with one install command, curl -fsSL https://aienos.com/install.sh | sh (www.aienos.com/alpha/). The GPU path is qualified only on the DGX Spark GB10; the three-machine physical qualification has not run yet.
• AIEN Rig for RTX gaming PCs: engine in a 12 GiB budget, the rig-tune frame-rate tuner, speech to text, a creator helper and a Windows app; feature-branch Windows CI passed on 2026-10-19, while aien-rig/main remains plans and placeholders. Nothing has run on the 5070 yet, 0 releases.
• Aienux on the iMac: the open pull request stack (netboot, keyboard, boot disk, screensaver), the goodbye animation, and Allen talking to the 27B through a UDP courier in the kernel (host tests pass, QEMU end to end next).
• Aienux on the Pi 5: ALLEN REFLEX decision making on the native Pi image (PR #402) and the Pi to Spark task protocol (PST1 draft).
• Research programme: an offline cost simulator (omega #583), a bounded learning-mechanism search (omega #584), the plan-reuse payback campaign (architecture #276), and a published verdict on a symbolic rewrite layer (VERDICT.md: cut 2 of 7 action graphs, no measurable cut on 27B decode).
• Public record: the Spark becomes the sole test judge with the hosted check only confirming the bundle (aienos PR #397), signed receipts in aien-receipts, every agent chat in agent-archive, and uncommitted agent work snapshotted to wip branches so no measurement stays on one disk.

AienOS is the reserved home for the core tree; the code still lives in the component repositories below and will be imported one at a time with full history. Repository names end in -mv today; the older names reach them only through GitHub's redirect.

Archived repositories are not listed. Each directory in the core repositories carries a .crumb file describing its purpose; a change that leaves the crumbs stale fails its gate.

Papers are published at www.aienos.com/research/, each with a PDF and a SHA256SUMS file. None is peer reviewed.

The research status record (what is measured, what is blocked) is at www.aienos.com/research/status/.

Since the 2026-10-24 policy change, GitHub-hosted AIEN Checks on pull requests and main are advisory: they check crumbs and the enrolled repository's portable smoke gates. They do not download a signed Spark receipt, and green does not authorize a merge. The AIEN Alpha job runs on alpha/* tags only and requires a full signed Spark bundle for a tagged commit on main (architecture workflow). The former PR receipt bridge and required hosted check are historical, not today's merge gate. Maintainers review and decide whether a change merges; a hosted badge alone is insufficient.

The default branch of aien-rig has no build or workflow yet. Its 2026-10-19 Windows CI run was on tuner/v0, not main, and does not qualify an RTX 5070. open-humanity is not covered by a blanket Spark-judge claim.

To get involved: read the architecture repository first (its PLAN_AUTHORITY.md says which document wins when two disagree), and open issues on the component repository the change belongs to.

• Closest to the metal. The fastest correct path wins. Software stays out of the hardware's way, and headline numbers are reported as a fraction of the machine's physical limit.
• Sovereign. No outside dependencies in the trusted base (the kernel, Omega, the host runtime): no CUDA toolkit, no systemd, no Python there, and builds are offline and reproducible. Outside the trusted base today: aien-rig describes its own C++/CUDA engine for RTX cards, and the Pi house server runs Linux with systemd services.
• Rust is scaffolding, Omega is the destination. Rust carries the host runtime today; the long-term home for every component is Omega. C only where the hardware justifies it, assembly only where measured.
• Evidence or nothing. A result is not demonstrated until a receipt for it exists. A target is never evidence of itself.
• Compare only with ourselves. Progress is measured against the machine's physical ceiling and our own previous run, never against other people's systems.
• Open. Compiler, language and runtime repositories (omega-mv, CUDApp, omega-bridge, physics-mv) are public under Apache-2.0 WITH LLVM-exception; standalone protocols and benchmarks stay plain Apache-2.0. Forks keep their upstream licences. The model served today is Qwen3.8-27B under Qwen's own licence; no AIEN-trained weights exist yet.

Apache-2.0 WITH LLVM-exception for the compiler/language/runtime tree (LICENSE), plain Apache-2.0 for interplane-mv and benchmarks-mv, and AGPL for the repositories relicensed on 2026-10-04 (among them aien-architecture-mv, aegis-runtime-mv, allen, spark-rsi-mv and aien-rig), except where a repository's own LICENSE says otherwise; aien-protocols-mv specifications remain under the Community Specification License 1.0. See each repository's own LICENSE file.

_{AienOS is not a UNIX and makes no such claim. UNIX is a registered trademark of The Open Group. NVIDIA, DGX and Grace Blackwell are trademarks of NVIDIA Corporation. Raspberry Pi is a trademark of Raspberry Pi Ltd.}