A small GPT-style model, trained entirely from scratch (no pretrained weights, no hardcoded responses), with a real tool-calling loop (calculator, sandboxed code execution, files, and more). Every reply below comes from the model's own learned parameters; every tool result is a real result from actually running that tool. See Docs & benchmarks.
Small model, trained from scratch on a small dataset — expect short, sometimes rough or wrong replies. Tool use only activates when the model's context window has room for the tool specs (see honest limitations).