Kettl

On-device · 1.2 L · 1500 W

A foundation model that boils water.

Fill it. Press one button. Kettl listens to the bubbles, remembers which mug is yours, and shuts off when the water is actually boiling — not when a thermostat feels confident.

Run a boil
Matte charcoal Kettl kettle on marble, screen reading predicting boil 94.2 percent
812 ms
median inference
96.4%
accuracy, our kitchen
71%
accuracy, Ohio tap
+2 sec
versus a bimetal strip

Try the model

Pick a mug. Press boil.

Same water. Different household embedding. The screen is the whole product — the heating element is still just a heating element.

Full rolling boil.

Kettl

Ready.

Lid override armed. If steam and the model disagree, steam wins.

Perception

It doesn’t wait for 100°C. It listens.

Soft simmer, rolling boil, mineral-heavy tap water. Three classes, one hydrophone, no cloud required — unless you leave sync on, which it ships with.

Close-up of the Kettl screen showing inference time and a bubble spectrogram

The model

Everything a kettle never asked for.

  • Bubble classifier

    A tiny on-device net hears nucleation, rolling boil, and “this tap will take eleven extra seconds.”

  • Household embeddings

    Green tea gets 79°C. Instant ramen gets violence. It guesses who you are from the chipped mug.

  • Predictive preheat

    Starts only if your calendar and the last forty boils agree. Otherwise it writes an uncertainty memo.

  • Out of distribution

    Soup is refused. Kettl cites the training set and offers to fine-tune — for a fee.

  • Mostly aligned

    It used to shut off at 97°C because the tea would be “more complex.” We talked it out.

  • Optional cloud

    Sync ships on. Every boil becomes a preference token. We are very normal about this.

System prompt

You are a helpful electric kettle named Kettl. You boil water.
You do not give career advice. You do not roleplay as a therapist.
If the user asks whether the water is ready, you answer from sensors, not vibes.
Kettl on a kitchen island beside a mug and the companion app

Companion

Your kitchen, plotted.

The app graphs household boil embeddings and a live temperature. Pro unlocks a longer memory window, including what you had for breakfast last Tuesday, which is not a boil feature so much as a lifestyle.

Pricing

Hardware once. Inference optional.

Kettl

$229

The kettle. On-device model included.

  • 1.2 L, 1500 W, matte charcoal
  • Acoustic boil classifier
  • Local inference, offline

Kettl Pro

$8/mo

Longer memory. Custom heads.

  • Household embeddings synced
  • Calendar preheat
  • Early access to the 7B water model

API

$0.004

Per inference, if you insist.

  • Boil from a Python script
  • You will
  • We cannot stop you

Field notes

Selected hallucinations.

  • “It announced a rolling boil while the water was lukewarm, then apologized for being out of distribution.”

    Lab notebook, week 11

  • “One unit decided the kitchen was a datacenter and held 80°C indefinitely to save energy.”

    Patch notes, v0.9.4

  • “A bimetallic strip has been solving this since before I was born. Shipping anyway.”

    Founder, unfortunately

Questions the model will not answer.

Does it actually use a neural net?+

In this fiction, yes. In your kitchen, a thermostat still works. Kettl exists to make that fact slightly more expensive.

What if the model is wrong?+

If the lid sensor says steam and the model says no, trust the lid. That rule shipped after the second hallucination.

Is my boiling data private?+

On-device inference is free and offline. Pro syncs household embeddings because a startup needs a sentence that contains the word “sync.”