One model. Different modes.

Give a model access to your own documents.

Your files are chunked, embedded and stored in a local vector index. At query time the passages that actually matter are retrieved and placed in the prompt, so the model answers from your material instead of its training data — and can cite which document it used.

Teach a model a behaviour, format or domain style.

A LoRA or QLoRA adapter is trained on your instruction dataset. Instead of rewriting every weight, a small low-rank matrix is learned alongside the frozen base — which is why a 14B model can be customised on a single consumer GPU in an afternoon.

Make a model fit, and make it fast.

Quantization trades a measurable amount of quality for a large amount of memory. Performance mode sizes each option against your actual hardware, converts to the format your runtime wants, and shows what the trade costs before you commit to it.

Nothing leaves the machine.

Every stage — parsing, embedding, training, inference — runs against local compute. ModelLM labels each operation local or cloud before it runs, and Private mode refuses to start any operation that would send your content off the machine.

Give a local model tools and let it act.

Tool definitions are exposed to the model, calls are parsed from its output and executed under an explicit permission policy. The hard part is not the loop — it is picking a base model whose tool-call formatting is reliable enough to build on.

Build a local coding assistant.

Code models are selected on fill-in-the-middle support, repository-scale context and latency rather than general chat quality. Coding mode biases the catalogue toward those properties and configures the project for editor integration.

Shape voice, format and refusal behaviour.

A system prompt plus, where it earns its keep, a small style adapter. Personality mode starts with prompting because that is almost always where the answer is, and only escalates to training when prompting demonstrably fails to hold.