ldraw-nova: AI agents build LEGO models by writing Python geometry
In ldraw-nova, models write a plan and a Python generator instead of raw coordinates, but nothing checks physics or stability.
taiga-s1: a 1.2M-parameter model drives FreeCAD via valid actions
taiga-s1 builds FreeCAD parts at about 1 ms per decision because the app lists valid commands each step; check your app's interface first.
Huseyin Kaplan · 1 min read
A 1.2M-parameter model built complete CAD parts in FreeCAD this weekend. About 1 ms per decision on a CPU. No screenshots.
Shiv Shanmugam's taiga-s1 is the fast half of an agent. A planner writes an ordered feature list. The small model picks each command.
The part worth copying is the interface. At every step FreeCAD reports its live state and the commands that are valid right now. The model only scores that list. Numbers come from the goal.
Trained on goals of up to 5 features, it finished all 100 held-out 11-feature goals. The card credits randomized position IDs first.
Before you reach for a bigger model on a UI agent, check:
- Can the app report its state as data?
- Can it list the valid actions per step?
- Can a wrong step be undone?
Caveat: every test goal is synthetic.
A small model can drive an app that tells it what's possible.
