Bud Gaia
The Personal AI Operating System — six layers on one machine, running many models and agents at once on the hardware you own. Gaia is the device-side version of Bud Novaria: the same architecture, scaled from the datacenter to a desk, a shelf, or a vehicle.
An operating system for the box, not another model runner.
Desk-side machines now ship datacenter-class AI compute with a large unified memory pool. The software on them still runs one model at a time, by hand. Gaia is the missing layer: many models and agents resident together, sub-second switching, automatic model choice, and your own work always first in line.
Six layers. One box. One install.
Gaia sits between the host operating system and the AI applications above it — the layer that today's local-AI stack is missing. Each layer is addressable on its own; together they turn a single machine into an AI department.
Above the stack: whole AI applications — agents plus a real interface — and the people and devices around the box: laptops, phones, front desks, cameras, vehicles.
Below the stack: your host OS and your silicon — GB10-class desk-side systems, RTX workstations, Core Ultra and Ryzen AI laptops, Apple silicon. Layer 01 docks them without application changes.
Six capabilities that need an OS to exist.
What the overview page states, expanded to the specifics an evaluator needs. Each one requires control of the whole machine — none of them can be a feature of a single application.
One request, end to end — on one machine.
When a person or an agent asks Gaia for something, the request crosses all six layers on the same box, in one scheduling domain. There is no network boundary to pay for and nothing to serialize between steps.
The request lifecycle — eight steps, one machine
Brains and skills — the memory model
| Resident unit | Scale | Covers | Example skills attached |
|---|---|---|---|
| Fast brain | 3B · ~1.9 GB | Routine language work, classification, triage | email tone (14 MB) · ticket triage (11 MB) · routing (9 MB) |
| Deep brain | 8B · ~4.6 GB | Reasoning, long-form drafting, domain work | sales conversations (22 MB) · financial analysis (26 MB) · Malayalam (31 MB) |
| Senses | speech + vision · ~1.2 GB | Transcription, speech, documents, camera | your accent (7 MB) · your document layouts (12 MB) · meeting voices (9 MB) |
| Frontier (optional) | off-box | The hardest public tasks, on request | routed via Gateway, inside a spend cap |
Sizes are illustrative of the shape, not a fixed bundle: breadth comes from several brains, depth from many megabyte-scale skills — and skills trained on your own runs stay yours.
Agents are declared, not scripted
The file names needs, not models. The same agent runs on a GB10, a laptop and a Mac; discovery resolves the best available fit on each.
The promise is a scheduling input. The orchestrator admits work it can keep and surfaces what it cannot, rather than silently missing it.
Permissions are explicit and revocable; anything outside the grant asks before acting.
Discovery, step by step
| Candidate | Good at the job? | Fits the box? | Around a second? | Verdict |
|---|---|---|---|---|
| 13B generalist | Decent | 26.8 GB — won't fit | — | Rejected pre-download |
| 3B chat model | Weak on the domain | 6.1 GB | ~0.3 s | Kept as fallback |
| 7B domain-tuned model | Best in class | 14.2 GB | ~0.6 s | ✓ Selected |
Illustrative Bud Simulator output. The point is the order of operations: predict fit on this chip, then download — never the reverse.
One box, four shapes.
Gaia is the same install in every shape below; what changes is who the box serves. Nothing here requires a datacenter, a cloud account, or an internet connection.
A workstation or laptop serving one person: agents in the gaps, private drafting, local research — and the machine still feels like yours.
One box on a shelf serving every device in a clinic, café, studio, law firm or workshop. No accelerator on the clients, no per-seat meter.
A production line, remote school, ship or field office — offline by default, low-latency, sovereign, with frontier access optional.
Robots, drones and vehicles that need on-site reflexes: local brains, hard promises, fault isolation between models.
| Shape | Who it serves | Network | Frontier calls | Notes |
|---|---|---|---|---|
| Personal | 1 person | local only | Opt-in per request | ✓ Full stack, host OS unchanged |
| Office hub | 5–50 devices | LAN | Opt-in, budgeted | ✓ Clients need no accelerator |
| Edge site | site devices | LAN, intermittent WAN | Usually off | ✓ Offline-native |
| Physical AI | one machine | on-board | Off | ✓ Fault isolation per model |
Target hardware
The 2025–26 generation of desk-side AI machines: roughly a petaFLOP of AI compute and a large unified memory pool, at a one-time $3–4k.
Host operating systems
Gaia installs on top of the OS you already run. It schedules AI work; your desktop stays in charge of the desktop.
APIs & standards
One OpenAI-compatible endpoint fronts local models and frontier providers alike, so existing clients point at Gaia by swapping a base URL. Agents and skills are declarative files; the SDK declares needs, never filenames.
Every number, with its basis and its status.
Gaia is pre-release, so this section is deliberately conservative: what is a hardware fact, what is a published third-party figure, and what is a Bud design target we intend to be held to.
| Claim | Status | How it will be measured |
|---|---|---|
| Sub-second model switching | design target | Time-to-first-token from cold, per model class, on each reference machine, with a stated resident set |
| Several brains resident together | design target | Concurrent resident models and blended skills within a fixed unified-memory budget |
| Zero idle footprint | design invariant | Resident bytes and accelerator utilisation with no request in flight |
| Pre-emption in favour of the user | design invariant | Time for a foreground workload to reclaim memory and compute after it starts |
| Fit predicted before download | design target | Bud Simulator predicted vs observed memory and latency, per candidate and chip |
| Skills improve week over week | design target | Judge-scored task win-rate for ART-distilled skills against the base brain, per workflow |
| Same architecture as the enterprise tier | structural | Shared lineage with Bud Novaria's Layer Zero, virtualization, gateway and agent runtime |
Against the obvious alternatives.
Nothing available today does the OS job on a personal machine. Each alternative solves one part — and leaves the scheduling, discovery and sharing problems to the user.
| Alternative | What it gives you | What Gaia adds |
|---|---|---|
| Local model runners | One model, running offline, quickly | Many models and skills resident together, automatic choice, sub-second switching, fair sharing of the box |
| Frontier subscriptions | The best models, per seat, per token | Everyday work answered on-box at the price of electricity; frontier as an explicit, budgeted choice |
| Rented cloud GPUs | Elastic capacity for bursts | Hardware you own, data that stays in the building, and no meter running on routine work |
| DIY orchestration on a workstation | Full control, if you have the team | Discovery, admission control, pre-emption and fault isolation as OS services rather than a private project |
| Bud Novaria | The enterprise AI operating system, in the datacenter | The same architecture at personal scale — one machine, one person or one building, no control plane to run |
Where a personal AI OS lands hardest.
The pattern repeats: a small organisation with real work, private data, and no appetite for a per-seat AI bill — or a machine that has to think where the cloud cannot reach.
An AI staff for the price of power
A café, clinic or law firm runs sales, scheduling, intake and follow-up on one box — no per-seat subscription, no per-token meter, nothing metered as the team grows.
Files that must not leave
Client files, patient records and financials are processed in the building. Frontier calls are a deliberate, budgeted exception rather than the default path.
Ship an app, not an install guide
Declare the models and skills your app needs; Gaia resolves them per machine. One binary works on a GB10, a Core Ultra laptop and a Mac.
Where the cloud does not reach
Factory lines, remote schools and clinics, ships and field offices: offline-native, low-latency, sovereign — with the same apps as head office.
Software that earns the box
An AI workstation is only as valuable as what runs on it. Gaia is the layer that turns a spec sheet into applications a buyer keeps using on day thirty.
On-site reflexes
Robots, drones and vehicles get local brains with hard promises and fault isolation, so one overloaded model cannot take the machine down.
Both tiers of the same architecture.
This brief is the Gaia reference. The overview page carries the story and the stack explorer; Novaria is the enterprise tier of the same architecture.
Back to overview
Bud Gaia
The Personal AI Operating System
The story in full: the machine, the wall, the six-layer stack explorer, the hub, and the thesis — every platform waited for its operating system.
Back to the overviewPlatform · eight products
Bud Novaria
The AI Operating System for the enterprise
Eight natively integrated products on one control plane, from silicon to agents — the datacenter tier of the architecture Gaia runs on a desk.
Explore Bud NovariaPut your data on it.
The fastest way to see what an integrated AI operating system does for your enterprise is a proof-of-concept on your infrastructure, with your data.