Breaking

Shiv Shanmugam Releases Taiga-S1, a 1.23M-Parameter Model for FreeCAD Command Execution

October 1, 2026 at 14:18 EDT

AI Agents
Open Source

Shiv Shanmugam has released Taiga-S1, a 1.23M-parameter model that operates FreeCAD by choosing the next valid CAD command rather than generating text or reading screenshots. The release tests a split-agent pattern: a planner defines the part, while a tiny local executor handles the command-by-command workflow.

2026 · Taiga-S1

A tiny executor for real CAD commands

Taiga-S1 splits the CAD-agent loop: a planner defines the part, and a 1.23M-parameter local model chooses the next valid FreeCAD command in about one millisecond on CPU.

1.23M

parameters

~1 ms

per CPU decision

24k

synthetic sessions

590k

training decisions

Generalization beyond training length

5 features

training ceiling

11 features

100% built correctly

17 features

95% built correctly

How the loop is split

Planner defines ordered features

→

Taiga-S1 scores valid commands

→

FreeCAD executes the next step

Why it lands

Local CPU execution, no screenshots, no LLM call per action, and fast command choice for repetitive CAD steps.

Where it stops

It needs a structured goal, does not choose design intent, and remains scoped to FreeCAD PartDesign workflows.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.