Training models isn't just for the big labs
Divergent Studio turns Claude Code, Codex and OpenCode into a research lab on your Mac. Beat frontier models at your domain task, for a fraction of the cost.
Autoresearch, on your Mac
Give Studio a goal and your coding agent runs the research loop until it beats the model you run today.
- ReadIt reads the papersEvery change cites the paper it came from.
- BranchIt branches the experiment treeEach idea is a branch. The winners get children.
- MeasureIt measures every runEach run is scored against its parent, so the tree only climbs.
Why every company will train its own models
Specialists win
Companies were never run by one generalist. A model trained for one job beats a frontier model at it, for less.
Agents do the research
Coding agents accelerate the training loop themselves, trying many ideas in parallel, until you beat the frontier.
You own it
Your weights, your data. Nothing you train ends up in someone else's product.
What you can train
Classifiers
Route tickets, tag documents and flag risk, in your categories.
Extraction
Pull the fields you need out of forms, invoices and contracts.
Agents
Trained on your tools and your workflows, not someone else's.
On-device
Small enough to run in the browser or on a phone, with no API bill.