Today we are releasing BELLA-v1.0-13B, an open-source foundation model that generates fully parametric, editable CAD from text and image prompts. Weights, evaluation code, and the benchmark harness are available now.
What BELLA is
BELLA is a 13B-parameter vision-language model fully fine-tuned on our parametric CAD corpus. It does not output meshes or images — it outputs feature trees: sketches, constraints, extrusions, revolves, and patterns that open directly in standard CAD tools and remain editable.
Evaluations
On the same benchmark we used to evaluate GPT-5, BELLA-v1 substantially outperforms much larger general-purpose models.
What it can't build yet
- Large assemblies — multi-body models beyond ~20 parts degrade quickly.
- Freeform surfaces — BELLA is trained on prismatic, engineering-style geometry; lofts and organic surfacing are out of distribution.
- Tight tolerancing — dimensions are usually right, but GD&T-level intent is not yet modeled.
Why open source
Our LoRA experiments convinced us that real progress in 3D reasoning requires full fine-tuning, open weights, and honest evals the community can reproduce. BELLA is the first step: use it, break it, and tell us where the geometry fails.
Get the model
Weights and the eval harness are available on request while we finalize hosting — email [email protected] and we will send you the checkpoint.
