Your own AI model.
On the hardware
you already have.
We train a model for the one task you need done and run it on the laptops, servers and browsers you already own. Your data never leaves the building. The model is more accurate on your task than a general one, and it costs a fraction of an API call.
Don't take our word for it. Run it right here, on your own GPU.
Contact updates appear here.
Private by default
Where your data goes.
- Your customers' data leaves the building on every call
- The model changes whenever they ship
- You pay for a huge general model to do one narrow task
- The request never leaves the device
- The model is a file you own; it changes when you retrain it
- Small enough to run on hardware you already have
Trained, not prompted
More accurate on your task.
Same 3B base model, same prompt. The only difference is training on examples of the task.
Contact notes: the untrained model never used the required tool-call format, so we scored its facts only and ignored format.
And it costs less
Stop paying per call for repeated tasks.
$120,000 annual savings
How a micromodel gets made
How we built the jazz model in the demo.
- 1
Collect examples
It is Wednesday 5:00 PM. Tonight's shows: 1. Eri Yamamoto Duo @ Arthur's Tavern, 5:00 PM 2. Jazzy Wednesday @ The Parlour at the Barclay… … 9. The Cookers @ Smoke, 6:30 PM — Eddie Henderson… … 44 shows → 30, 9, 11, 27, 13
536 examples like this one, each labeled using a written rubric. The data never left the machine.
- 2
Train on a laptop
Training loss · Jazz top 5 Step Loss 10 1.299 20 1.029 30 0.856 40 0.831 50 0.774 60 0.802 70 0.705 80 0.778 90 0.713 100 0.627 110 0.591 120 0.608 130 0.589 140 0.506 150 0.45 160 0.538 170 0.472 180 0.439 190 0.487 200 0.458 210 0.422 220 0.415 230 0.437 240 0.425 250 0.412 260 0.36 268 0.4 The actual training run: 31 minutes on one MacBook. No cloud GPUs.
- 3
Deploy one file
jazz_top5.gguf29 MBIt runs on our own WebGPU runtime, which works on any OS, in Chrome and Safari, and inside your own apps. The demo above is that runtime, running in your browser.
Is your task a fit?
Good fits are narrow, repeated tasks.
- Label every support ticket
- Pull fields out of invoices and forms
- Route inbound texts and emails
- Turn call notes into CRM updates
- Flag documents for human review
- Rank a feed by your own taste
- Answer anything about anything
- Write open-ended long-form content
- A task you run twice a year
Already in the field
A document classifier for a tax firm.
- 5 days
- from raw data to a trained specialist
- 1 MacBook Pro
- on the firm's premises, no rented GPUs
- < 4 hours
- of human time; the rest ran unattended
- 0 bytes
- of client data left the building
Our team's track record
- Coding agent in production, two weeks after GPT-4Devin: Mar 2024 · Claude Code: Feb 2025
- Agent skill files for teaching agents repeatable tasksAnthropic Agent Skills: Oct 2025
- Retrieval-augmented generation in productionA default enterprise feature: 2024
Built where data security isn't optional
- Financial servicesIntegrations with 18 financial institutions
- TaxThe AI platform at a tax firm with 1,700 clients
- Children's gamesA game studio built to $1M in revenue on Roblox
FAQ
Common questions
Where does our data go?
Nowhere. Training and inference run on hardware you control, and the runtime doesn't phone home. We never see your training data.
What hardware do we need?
Training runs on a Mac today, with other operating systems and GPU types rolling out soon. The models you make run on what you already have: macOS, Windows, Linux, iOS and Android, in Chrome and Safari, and inside your own apps. The demo above is running on your device right now.
Who owns the model?
You do. It's a file on your hardware. It doesn't change unless you retrain it.
What happens when it's wrong?
You correct the output and retrain; that refine loop is how every specialist gets built. Each one also sits behind a scope gate that refuses off-topic requests in milliseconds, before the model runs.
What does working with us look like?
We sit with your team and train the first model hands-on, on your hardware, with your data. Our first engagement went from raw data to a trained model in five days.
What task do you want to automate?
We'll tell you if a micromodel can do it. Or email hello@micromodel.company.