ZeroState Labs is a research and product studio for local AI. We build systems that think, remember and act on the hardware in your home, never in someone else's data centre, and we treat that constraint as the source of better products rather than a limitation. Alongside the lab sits ZeroState Press, our publishing house for books written with AI.
For ten years, "AI" has meant a very large model in a very large building, reached over a wire. That era produced remarkable tools and a quiet bargain: to be helped, you had to be heard by someone else. We think that bargain is ending.
Two things changed at once. Small models became good at the jobs that matter most in daily life: listening, remembering, keeping company, recognising a voice, labelling what's in a room. And the boards that can run them, cheaply and at conversational speed, arrived in the form of neural accelerators on sixty dollar computers and phone-grade chips.
Put those together and a new kind of product becomes possible: one that is more private, faster, cheaper to run, and more trustworthy than its cloud equivalent, precisely because it never leaves the room. ZeroState Labs exists to find those products, prove them, and ship them.
Our mandate is broad on purpose. Companionship, robotics, agentic systems, new hardware: anywhere a local model can do a job better than a remote one, we want to be working. The projects below are the first extensions of that mandate, not the limits of it.
Every system runs fully offline in normal operation. Cloud fallback may exist, and is off by default.
No camera on a companion. No frame or clip that leaves the home network. Nothing collected that the owner can't see.
The people we build for want to be remembered and kept company, not out-reasoned. A consistent voice with real memories is the product.
A model with a smaller range of capability has a smaller range of ways to go wrong. That matters everywhere and most with children.
Systems that act on their own answer to a transparent local record and a human who can see it. Autonomy without a hidden log.
Prove the emotional or behavioural hypothesis on a phone before spending on hardware. Retire risk early, then build the body.
Each area is a place where running locally is not a compromise but the reason the product works at all.
Small on-device models with long-lived episodic and semantic memory, a loadable persona engine, and tone that adapts to the person. For older adults living alone, and for children, with the controls each audience actually needs.
Groups of small, cheap robots with a shared local brain: role assignment, a common map, knowledge that spreads between units by talking. Interesting at three, still cheap at thirty, no server bill at any number.
Agents that read, decide and act on local hardware: answering a robot's mail in character, running a household's routines, drafting for a human to approve. Grounded in a local event log, never a black box.
Raspberry Pi 5, RK3588-class boards, ESP32-S3 modules and phone-grade NPUs, with inference engines such as Cactus and llama.cpp targeting them directly. We port early, measure honestly, and design form factors around what the silicon can do.
Ideas that only make sense when the model is in the room: a toy that knows your child's voice, a beep language robots and people share, a story read aloud with a bookmark the model can't lose. We prototype fast and keep the ones that earn it.
Transparency dashboards, on-device parental and caregiver controls, encryption at rest, export and delete by the owner, and the regulatory work (COPPA, GDPR-K, toy safety) that lets these products reach real homes.
Local inference, memory, a persona engine and a transparency surface are shared. A new project is mostly a new personality, a new body, and a new audience, which is why we can run several at once.
A conversation and memory core, or a behaviour simulator, running on a laptop or phone. The emotional hypothesis gets tested before anything is soldered.
The same code on a sixty dollar board. If the experience survives that, it can live in a home affordably.
A plush, a tabletop device, a 48 mm cube. Form factor comes last, shaped by what the hardware proved it could do.
Families, classrooms, care settings. Certification and privacy review before any pilot touches real data.
$ZERO is a utility token on Solana that funds ZeroState's development pipeline — the shared local-AI hub behind Companion, Hearth and Hoplings — transparently, with the community that wants them to exist. No presale, fixed supply, locked liquidity and a public treasury. It hasn't launched yet: try the demo purchase and read the draft white paper.
Each project is an extension of the studio's work into one audience. They share the core, and each one has its own page and white paper.
A private, always-on-device presence that listens, remembers who you are and what you told it last time, and encourages family contact without replacing it.
Four original in-house characters, Skye, the Timberfolk, Fenn and Nightlight, each with its own memory of a child, and a parental control system built in from the start.
48 mm cubes with expressive faces and a local hub that runs the models, keeps the shared map and directs the group. They learn the room from each other and answer their own mail in character.
Watches the crib, listens for cries, tracks the room, and works with the vital-sign wearables families already own — with every frame of video and every heartbeat processed and stored inside the house.
A team of AI analysts reads live market data, behind a risk wall written in code, and every decision is logged. Whether that gives an edge is unproven, so the first thing being built is the apparatus that tests it.
Not investment advice. Paper trading only. No returns are promised or implied.
ZeroState Press is our publishing house for AI-written fiction and non-fiction. It's where the same curiosity that drives the lab gets pointed at story: what happens when a model is treated as a writing collaborator with a real editor, a real house style, and a real reader at the end of the line.
The Press runs on the same conviction as the lab. Tools should belong to the people using them, the craft still matters, and the interesting work is in the seam between what a model can produce and what a human decides to keep. Every title goes through a human editorial pass, a house typesetting pipeline built for e-ink readers and print, and honest labelling about how it was made.
We're looking for pilot partners in eldercare and education, hardware collaborators on embedded ports, researchers working on small models and edge inference, and investors who think local AI is where the next decade of consumer products lives.
Sign in to set up a profile — who you are and what role you're coming from — and request access to Hoplings, Companion, or Hearth. We review every request by hand.