- Bud Gaia is the Personal AI Operating System — the device-side Bud Novaria, running on the supercomputer on your desk.
- The problem: the hardware already exists — roughly one petaFLOP, 128 GB unified memory, $3–4k once — but it runs one model at a time, loaded by hand, so real work returns to the cloud.
- The solution: six layers — Experience, Agents, Access, Discovery, Intelligence, Metal — installed beneath Windows, Ubuntu or macOS.
- What it feels like: Alt-Tab for models. Roughly one second to wake (design target), zero GB idle footprint, and your work pre-empts the AI.
- The business value: no per-seat subscription and no per-token meter; five to fifty devices on one box; the data stays inside the building.
- The thesis: every platform waited for its operating system. The hardware is on the desk.
Bud Gaia. The Personal AI Operating System.
The device-side version of Bud Novaria — the same architecture, running on the supercomputer on your desk. Gaia does for models and agents what an operating system did for programs. Gaia is the bounded Earth; Novaria the unbounded realm above.
The story in 45 seconds ↓
Great demos. No day jobs.
One model at a time, loaded by hand.
This machine is not waiting for a better model. It is waiting for its operating system.
Six layers. One install. Your computer, still yours.
Installs on Windows, Ubuntu or macOS. Runs beneath everything else.
Alt-Tab, for models.
No request, no model in memory.
Your game, render or call reclaims the machine first.
Work attaches to the gaps you leave behind. Illustrative.
You outrank the AI. Always.
An AI staff for the price of power.
Nothing metered as the team grows.
Laptops, phones, tablets, front desk, cameras — no accelerator on the clients.
Sensitive work is answered on-box or refused; spend caps make a runaway bill structurally impossible.
From one great demo at a time — to a building full of AI applications, on hardware you already own.
The hardware is on the desk.
The operating system comes next.
Bud · Simplifying Intelligence.
Full agents, real products — not chat windows.
The AI that earns money and saves hours isn't a chat box. It's a complete application: agents plus a real interface, quietly using many models underneath. They already exist, in the open — and each one needs its own cast.
An AI sales rep: answers inbound, qualifies, books, follows up on every lead — by chat and by phone.
Reads every résumé in the pile, screens against the role, ranks and explains the shortlist end to end.
An open investment-research terminal with AI analysis over market data, filings and your own notes.
A modern open-source CRM with AI woven through the workflow: enrichment, summaries, next best action.
Workplace AI across a company's knowledge and tools — search, answer and act over everything you already have.
Open driver assistance operating real cars: perception and control loops that cannot wait for a network.
Third-party open-source projects, named for illustration; no affiliation or endorsement implied. Also in the roster: Openwork's back-office agent workforce — and a long tail behind it.
Every one of these is agents + interface + a cast of models. That is the unit of value — the "app" of the AI era — and today's local stack can run exactly one of them at a time.
Why almost nobody runs them locally.
Take one of those applications and look at what it actually needs to do its job. Then look at what today's local-AI software can give it.
On today's software stack, high-value local applications are not hard. They are impossible. The hardware is not the bottleneck — the missing layer is.
We have seen this film before.
In the DOS era a PC ran one program at a time. You quit WordPerfect to open Lotus, every application fought over memory by hand, and one crash took the whole machine down.
| The personal computer, 1985 | The AI workstation, today |
|---|---|
| One program at a time | One model at a time |
| Quit an application to switch | Unload and reload weights to switch models |
| Applications juggle memory themselves | Every application fights for VRAM alone |
| Hunt for a driver on a floppy disk | Hunt for a model on a hub |
| One crash freezes everything | One overload takes down the box |
Six layers, one personal AI operating system.
Gaia installs on top of Windows, Ubuntu or macOS and runs beneath everything else: many models and agents at once, instant switching, fair sharing of the machine — and it yields to your own work first. Select a layer.
What you touch
AI applications install like phone apps, with settings, health and permissions in one place — and an SDK so developers ship an app, not an installation guide.
- App Store: install a full agent-plus-interface application in one step
- Permissions like a phone: each app gets exactly the access you grant
- One health view: what is resident, what is running, what it cost you
- SDK: declare the models and skills an app needs, never a filename
How work gets done
Agents are first-class citizens of the OS — with tools, memory, schedules and a service promise the scheduler is accountable for. And they train new skills from their own work.
- Declarative agents: a small file describes needs, schedule and permissions
- Promises, not hopes: "the brief is ready by 07:05" is scheduled, not attempted
- They live on the box: close the laptop and the agents keep working
- ART: agentic reinforcement training turns your workflows into new skills
One door to all AI
A single standard API fronts local models and frontier clouds alike. Gaia routes by privacy, difficulty and cost — and is always explicit about what ran where.
- Private by default: sensitive work is answered on-box, or not at all
- Frontier by choice: OpenAI, Anthropic, Gemini and ElevenLabs behind one API
- Zero-code swaps: change the model behind a call without touching the app
- Spend caps: a budget that refuses a runaway bill
Which model, automatically
Choosing a model is the system's job, the way finding a printer driver is. Applications ask for an outcome; Gaia finds the candidate that can deliver it on your exact chip.
- Model Finder: continuously ranks open models against real jobs, languages and licences
- Bud Simulator: predicts memory and speed on your hardware before a byte is downloaded
- Fit-aware: the shortlist accounts for everything already running
- Upgrades offered, not forced: when a better model ships, Gaia re-checks
The engine room
Several base brains stay resident, each carrying a stack of featherweight skills. Cold-start acceleration wakes anything else in about a second; idle costs nothing.
- Orchestrator: decides what is resident, what sleeps, what is pre-warmed
- Serverless: scale to zero — a box full of AI apps idles like a box with none
- Engine Backend: each model matched to the runtime that serves it best
- Skills, blended: several adapters live on one brain at the same time
Any hardware
One accelerator, safely shared: partitioned so a crash never spreads, and open to whatever silicon you own or add later. The same architecture that runs Novaria in the datacenter.
- FCSP virtualization: fine-grained partitioning of the whole device
- Layer Zero: one execution surface across GPU, CPU, NPU and friends
- Fault isolation: one overloaded model no longer takes down the box
- Add devices freely: extra hardware joins without application changes
Alt-Tab, for models.
With virtualization and cold-start acceleration underneath, a model that isn't even loaded is about a second away — and many run side by side, each application bringing its own cast.
Many models, one box
Chat, speech, vision and embedding models coexist. No evictions, no juggling, no restart to try something else.
Every app, its own cast
One app's sales models and another's research models live together and swap in around a second.
Nothing to manage
No loading, killing or reloading by hand. The OS decides what is resident — always.
It gets out of the way.
An AI OS that slows your machine down is a tax, not a tool. Gaia holds nothing when nothing is asked of it — and the moment you need the machine for a game, a render, a build or a call, it gives everything back.
Scale to zero
No request, no model in memory. A box full of AI apps idles like a box with none.
Pre-emption
Your game, render or call reclaims the machine at any moment. Gaia steps back first.
Opportunism
Lunch, evenings, the small hours: agents and overnight learning use the gaps.
Many brains. Many skills.
Small models are specialists, so Gaia never bets on one. Several base brains stay resident — language, reasoning, speech, vision — and each carries a stack of featherweight skills. A brain is gigabytes; a skill is megabytes. Several can be live on one brain at once, blended per request.
Overnight, on the idle accelerator, agentic reinforcement training turns your own workflows into new adapters. Breadth from many brains, depth from many skills — and the skill belongs to you.
You will never pick a model again.
Which open model is actually good at sales calls? Which fits on your box beside four other applications? Which keeps a one-second promise? Today that is a week of guesswork and 20-gigabyte downloads.
Ask for the job
"Good at sales conversations, speaks Malayalam, answers in about a second." An outcome — never a filename.
Search the world
Gaia continuously ranks open models against real jobs: skill benchmarks, languages, licence, size. It shortlists the ones that can do this.
Prove the fit first
Before a byte is downloaded, it predicts memory and speed for each candidate on your exact chip, beside everything already running.
Install, tune, keep watch
The winner is fetched, matched to the right engine, warmed and served. When a better model ships, Gaia offers the upgrade.
| Candidate | Good at the job? | Fits your box? | Around a second? |
|---|---|---|---|
| A 13B generalist | Decent | ✕ 26.8 GB — won't fit | — |
| A 3B chat model | ✕ weak on sales | ✓ 6.1 GB | ✓ ~0.3 s |
| A 7B sales-tuned model | ✓ best in class | ✓ 14.2 GB | ✓ ~0.6 s |
Illustrative shortlist — Bud Simulator's verdict is computed before downloading a single byte.
Private by default. Frontier by choice.
One standard API fronts everything — your local models and the frontier clouds. Gaia routes each request by privacy, difficulty and cost, and is honest about what ran where.
Agents you can rely on.
Agents on Gaia are not scripts you babysit. They are portable, declarative and accountable: describe one in a small file and it runs on any Gaia box.
Declare needs, not models
The discovery layer resolves the best fit on each machine — the same file, the right model everywhere.
Promises, kept
Every agent carries a service promise, and the OS schedules the machine to keep it.
Permissions like a phone
Agents get exactly the access you grant. Anything sensitive asks first.
They live on the box
Close the lid, walk away — the agents keep working. Which raises the question: who else can use them?
One day, four agents, one box
Illustrative day on a single box. Agents are scheduled against their promises, not run on a timer you maintain.
One box. Everybody's AI.
An AI workstation is not a personal toy — it is infrastructure for a whole building. Put one on the shelf and every laptop, phone, till and camera around it gets the same applications, agents and models. Nothing on the client but a screen.
Clients need no GPU
A phone or a five-year-old laptop gets the same AI as the box. The intelligence is on the shelf, not in your hand.
One box, one bill
Buy the hardware once and serve the whole team. No per-seat AI subscription, no per-token meter.
Agents don't sleep
They run on the hub, so they keep working through closed laptops, lunch breaks and weekends.
The data stays inside
Requests never leave the network unless you allow a frontier call. The building's data stays in the building.
This is how a five-person clinic or a fifty-person factory actually gets AI: one device on a shelf, everyone served, nothing metered — the office router, but for intelligence.
The same applications — now they run.
Back to those applications. On Gaia, the demo machine becomes an AI department.
Every platform waited for its operating system.
The hardware is on the desk. Novaria for the datacenter, Gaia for the desk — one ecosystem, both tiers.
The full argument, in depth.
The product reference behind this page — and the enterprise tier it's the personal version of.
Bud Gaia Product Brief
Six layers, the request lifecycle on a personal box, hardware and deployment shapes, and every claim paired with how it will be measured.
Read the product brief →Bud Novaria
The AI Operating System for the enterprise — eight products, one control plane, from silicon to agents.
Explore Bud Novaria →Put 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.