Post-training

Trains a small open model on an agent’s own production traces, then serves it.

Belvedir's docs say how many traces a training run needs and how a result is scored against the base model, so its claim that a small tuned model can replace a large one can be tested on traces we generate ourselves.

Turns your examples or traffic into small task-specific models you own.

Distil Labs gives two free training runs and a command-line tool, and publishes its benchmark and example repos, so its claim that a small tuned model can match a big one can be tested on a task of our own.

Free open-source software to fine-tune and run open models on your own machine.

Unsloth publishes its code and a free desktop app, so its speed and memory claims for fine-tuning can be checked on an ordinary home GPU.