A tiny brain.
Big pet energy.

PetInst-LLM turns everyday owner language into one safe, high-level pet intent—locally, quietly, and without taking control away from the app.

See how it thinks Open the lab notes

hears meaning

picks one intent

A cheerful cartoon virtual pet with a colorful, visible pet brain

built for
belly rubs,
not spreadsheets

scroll to explore

ONE INTENT AT A TIMELOCAL BY DESIGNTHE HOST STAYS IN CONTROLONE INTENT AT A TIME

Not a chatbot
in a fur suit.

PetInst is a specialist inside a larger, safer pet system. It understands the moment. Trusted code decides what actually changes.

Owner speaks

Finalized voice or an allowlisted activity becomes compact context.

PetInst understands

The tiny model selects exactly one eligible semantic function.

Host decides

Deterministic policy validates the call, then safely updates the pet.

The important boundary: PetInst never controls money, inventory, coordinates, files, or arbitrary code.

The harness keeps
the brain pet-sized.

Each moment gives the brain only three to five safe actions. It chooses one; anything outside that menu is ignored.

Builder detail

The answer must be one empty-argument semantic function call. The host validates it before any state can change.

pet_eatpet_restpet_playpet_seek_attention

See all 12 safe intents

pet_comfort_ownerpet_celebratepet_acknowledgepet_rememberpet_self_carepet_refusepet_withdrawpet_ask_for_help

“I had a really hard day.”

one safe choicepet_comfort_owner()

No generated chat. No secret commands. Just one bounded intent for the host to validate.

One brain.
A whole cast.

PetInst is not tied to one mascot. Give the same intent model a different character context—personality, needs, relationship, and memories—and each pet can still feel unmistakably its own.

Character changesvoice · needs · memories

Safety stayseligible intents · trusted host

Example personalities — one safety contract, different character context

Mori
curiousPip
gentleRex
boldLumi
bright

The architecture is designed for a cast. Today’s validated product slice still starts with one pet; multi-pet rollout remains a product step.

character context

shared intent layer

Four different original cartoon pets gathered around a shared glowing brain

Build your own
pet system.

Use PetInst as the small intent layer inside your own companion. You design the creature and its world; the model reads the moment; your trusted host decides what can happen.

Try it in the browser Get the model

01Describe the pet

Personality, needs, relationship, memories, and the current moment.

02Offer safe actions

Expose only the small set of semantic functions that make sense now.

03Let the host decide

Validate the call and keep state, rewards, purchases, and side effects in trusted code.

A small PetInst brain connecting locally to pet experiences on a phone, laptop, home hub, and wearable edge device

Live demoBrowserWASM or WebGPU

Integration targetPhone or tabletNative local runtime

Integration targetDesktop companionLocal GGUF runtime

Research targetEmbedded petDevice-specific runtime

Portable model, target-specific proof. Each device still needs its own runtime, memory, thermal, offline, interruption, and lifecycle validation.

Research,
with the messy bits in.

We did not ask a giant model to pretend it was small. We trained, challenged, rejected, retrained, quantized, and measured a compact specialist.

A warm cartoon research lab testing a small pet brain with speech examples and evaluation jars

selected
step 300

Teach

16,240 training examples pair everyday companion language with a safe semantic intent.

Challenge

Hard contrasts test the difference between similar phrases: comfort vs. attention, rest vs. self care.

Measure

Frozen balanced and natural-language controls check accuracy, strict calls, safety, memory, and ordering.

Package

The selected checkpoint is fused and quantized into a 278 MiB Q8_0 GGUF for local integration.

Selected research candidate

FunctionGemma 270M
Intent v2.2 · Q8_0

Desktop evaluation · not physical-device release proof

Paper link pending

94.16%balanced intent accuracy1,352 rows · Control G

93.70%natural-language accuracy1,000 rows · Control H

100%strict calls on balanced setexactly one valid tool call

278 MiBlocal Q8_0 artifact291,557,760 bytes

What is still open: release approval and physical-device verification. Desktop and Simulator evidence do not prove real-device memory pressure, thermal behavior, microphone flow, or airplane-mode readiness.

Small enough to live with your pet.

Safe enough to
earn its trust.

Try the pet system View the model card

A tiny brain.Big pet energy.

Not a chatbotin a fur suit.