A small model. A big mind.
Axom is an architecture that turns a small, unmodified American model into an assistant that researches, codes, remembers, and works on its own, right on your hardware.
We don't need bigger models. We need better architecture.
Axom doesn't retrain the model underneath it. It builds everything around it (memory, tools, skills, and judgment) so a model that fits on a desk does work that used to need a data center.
One assistant. Four jobs, done properly.
Researches
Searches the web, reads every page whole, and tells you which source said what. Pages carry their publish date, so last year's news never passes for this week's.
Codes
Reads, edits, and runs your project, with several tools at once when the work allows. It asks before it writes a file or runs a command, and remembers your answer per project.
Remembers
Every conversation, across every session. It notices things about you and proposes them as notes. You accept or reject each one, and see exactly where it came from.
Works on its own
Give it a goal and it plans, works, and reports back. Ask it to keep watch and it decides how often to check. Jobs survive a restart.
Never think about context again. It's a scratchpad.
Other assistants drag the whole conversation along until the window overflows. Axom rebuilds its context from memory on every turn: only what this turn needs, cited and sized to fit. The memory itself keeps everything, and never deletes.
Who you are
The notes you've accepted and how you like to be spoken to. Present on every turn.
This conversation
What you just said, kept whole. Its own long replies are trimmed so your words always fit.
Everything else
Searched from the full history by meaning and by keyword, then ranked, so a detail from months ago surfaces when it matters.
A page too long for the window isn't cut off. Axom reads it whole and keeps what answers your question.
Ten pages. Ten agents. One pass.
When research brings back more than fits, each page goes to its own agent and the whole swarm fires together. Their findings come back as latent positions, not text: up to 70 times smaller, and more accurate than summarizing.
The model is American. The mind is yours.
Underneath is Gemma 4: American-made, unmodified. On top sit leaves: small adapters that snap onto the model and give it skills, habits, and character.
Habits, built in
Axom's own leaf teaches the model how to work: search before answering, cite what it read, use tools cleanly.
Its voice, your call
Dozens of dials for how Axom talks: tone, length, humor, format. It watches how you write and suggests settings. You decide.
Rated, kept, exportable
Every turn is kept on your machine with your rating of it. That's the raw material for leaves shaped by how you work.
The first AI that grows with you.
Use Axom as your sysadmin and, at first, it makes sysadmin mistakes, because the chassis isn't a sysadmin. Every one of those mistakes is caught, fixed, and kept. Enough of them become a SysAdmin leaf, pressed from your own work. Axom stops making them, and the skill is yours.
Its mistakes become its training.
After every turn, Axom checks its own work against the record: a command that never ran, a file it said it saved, a report two sections short of what you asked. It shows you its read, quietly, beside the good and bad buttons.
Axom's read: probably wrong. Said it restarted nginx; it never ran. Agree · Actually fine
While you're away, a teacher replays each mistake on a copy of your workspace, knowing what went wrong, until it gets it right. A fix is kept only if it passes the same check that caught the mistake. No check, no row.
A mistake · a verified fix · a pile that becomes a leaf
Shows its work. Checks its claims.
Watch it think
Every search, read, and command appears as it happens. Every status line on screen is tied to something that actually ran.
No invented citations
Axom only says "according to" about a page it read this turn. Ask what's new and it goes and checks instead of answering from memory.
Interrupt anytime
Type while it works and it reads your note at the next step, even mid-answer, and changes course.
Small model. Big-model results.
Five X1 tiers, from 5 GB of GPU memory to 20. See the lineup.