CAD Assistance

1. No mature open-weight AI model exists specifically for CAD assistance yet (as of mid-2026). The field is nascent.
2. For parametric CAD assistance: use general coding models to generate OpenSCAD scripts. ollama pull qwen2.5-coder:14b → prompt: "Write an OpenSCAD script for a parametric bolt: M8 thread, 30mm length, hex head."
3. For FreeCAD automation: FreeCAD has a Python API. Use coding models to generate Python scripts that construct CAD geometry: "Write a FreeCAD Python script that creates a rectangular duct with 100mm×50mm cross-section and 200mm length with 3mm wall thickness."
4. For generative design (topology optimization): use code-generation models to produce FEA (finite element analysis) configurations for CalculiX or Elmer solvers.
5. For mesh-to-CAD conversion: no reliable open-weight model. Commercial solutions (nTopology, Materialise) dominate this space.
6. Reality check: CAD assistance with local AI in 2026 is "generate code for CAD programs" not "AI designs the part." The models understand geometry concepts but can't produce production-ready CAD files directly.

CAD assistance doesn't need powerful hardware. Coding models (Qwen 2.5 Coder 7B) that generate OpenSCAD/FreeCAD scripts run on any $300 laptop CPU at 20-40 tok/s. The CAD software (FreeCAD, OpenSCAD, Blender) runs on integrated graphics. A $400 laptop handles the entire AI-assisted CAD workflow. The bottleneck is whether the AI can generate correct CAD geometry — currently, simple parametric parts (bolts, brackets, enclosures) work; complex assemblies with constraints and tolerances do not. $400 gets you basic script generation for parametric CAD. $400 does NOT get you AI-driven generative design.

For CAD assistance specifically, hardware is not the bottleneck — software maturity is. Any RTX GPU (3060, 3090, 4090) provides trivial speedup for what is essentially a text-generation task (generating CAD scripts). A workstation with RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb) + Ryzen 7 7700X + 64 GB DDR5 + 2TB NVMe. Total: ~$900-1,200. This runs both the coding model and the CAD software (SolidWorks, Fusion 360 via VM, or FreeCAD natively). The $1,500-2,000 budget for "serious CAD AI" buys a workstation that also handles simulation (FEA/CFD) and rendering — the AI component is the cheapest part of the setup.

The mistake: Prompting a coding model with "Design a drone frame in CAD" and expecting a .STEP or .SLDPRT file as output. Why it fails: LLMs generate text. A CAD file is a complex binary format (or a structured text format like STEP) with hundreds of features, constraints, material properties, and tolerances. No model can directly output a valid CAD file — the token budget alone for a moderately complex part (extrusions, fillets, chamfers, mates) exceeds standard context windows. The fix: The model generates CAD Python scripts (OpenSCAD, FreeCAD API, CadQuery). You run the script in the CAD program, which produces the file. Prompt: "Write a CadQuery Python script that creates a drone frame: 250mm diagonal, 5mm carbon fiber plate, motor mounts at each corner with M3 holes." The script generates the geometry deterministically. AI writes the script; the CAD kernel builds the geometry. Text-out, geometry-out is the only viable path right now.