Taiga-S1: a 1.2M-param model builds complete CAD parts in FreeCAD — ~1ms per decision on CPU, no LLM, no screenshots. Your agent plans, Taiga executes
Developer Shiv Shanmugam (@shhivv) released Taiga-S1: a 1.2M-param transformer trained from scratch as the fast "System 1" execution layer for CAD agents — an LLM agent planner decides what to do, Taiga decides the next click each step. Given an ordered feature goal (e.g. "plate 40x30x10, Ø6 hole at (10,0), polar pattern x6, fillet top edges"), it reads FreeCAD's live state (feature tree, selection, sketch constraints) and picks from available commands: select plane, sketch, draw, constrain, pad, pattern, fillet. No LLM, no vision, no screenshots; ~1ms per decision on CPU, driving the real FreeCAD GUI over a local socket. Generalizes beyond training length: trained only on goals up to 5 features, it nailed all 100 held-out 11-feature goals (~55 commands) and 95% of 17-feature ones; 99.8% per-step accuracy vs the teacher (IoU>=0.99). With 20% of actions replaced by random ones it notices, undoes the damage and finishes 86-100% of parts. Keys to generalization: randomized position IDs (Ruoss 2023), coupled ordinals between goal items and tree features, a modular "done?" policy, factorized feature types. Trained on 24k synthetic headless-FreeCAD sessions (~590k decisions) with supervised training + two rounds of DAgger. Weights: huggingface.co/shhivv/taiga-s1; next step: accessibility-tree interface for apps without scripting APIs.
