Curate the data. Train it on your GPU. Watch it call your tools.

SLM is a workbench for models you own end to end: the corpus, the dataset, the weights, the tools they are allowed to touch. It runs on the machine under your desk, prints a citation for every answer, and shows you exactly who decided each tool call.

Three verbs, one machine

Curate: point a recipe at your folders, name what is excluded and why, review two hundred samples before anything trains. Train: from scratch on your corpus, or continue a bundle you already have, with the loss on screen. Call: ask the bundle, get an answer with citations, and see the tool call, its validated arguments, the confirmation tier and the result.

Nothing leaves the machine

The corpus is yours, the weights are yours, the tools are fenced to the roots you allow. There is no upload step in the pipeline. The only thing this site ever receives is a recipe you choose to keep in your account, and it never receives your files.

Honest about the model

The scratch model is about a million parameters: it learns your corpus and answers grounded in retrieval, and its tool calls are routed by a rule planner, not emitted by the model. The trace says so. A model that emits its own calls is the fine-tune lane, which is on the roadmap and labelled that way here.

What it costs

Downloading, the docs, and running both lanes on your own hardware are free and will stay free. The plan buys compute you did not run Python for: hosted fine-tune minutes and served bundles. That lane is not live yet, so the page says notify me, not buy.