A 4B‑Parameter Model Beats Top AI 3D Generators in Eight Benchmarks

Current AI‑driven 3D generation can produce visually realistic objects such as earphone cases or coffee cups from a single text prompt, but the output lacks the engineering parameters, constraints, and construction logic required for real‑world manufacturing.

Arko‑T, released by BitUnlimited, is a 4 billion‑parameter Transformer that performs Text‑to‑Structured‑3D generation. It accepts a natural‑language description of a single part and outputs a Build123d program that defines named parameters, structural features, and the full modeling process, which can be executed by a CAD kernel to produce a fully editable design entity.

In a direct head‑to‑head evaluation against seven state‑of‑the‑art generalist models—GPT‑5.2, Claude‑4.5‑sonnet, Gemini‑3.5‑flash, DeepSeek‑V4‑Pro, Qwen‑3.6‑max, Kimi‑k2.6, and GLM‑5.1—Arko‑T achieved the best score on eight of twelve quantitative metrics and ranked second on the remaining three, demonstrating a clear advantage in structured 3D generation.

Beyond accuracy, Arko‑T is highly efficient: on a local GPU it processes a single sample in an average of 0.41 seconds, costing only $0.28 per inference. This is at least an order of magnitude faster and up to 65 times cheaper than the cheapest baseline (DeepSeek at $1.69) and the most expensive baseline (Claude at $18.14).

The model also generalizes to components absent from its training set by leveraging learned structural knowledge to compose and generate novel parts.

Integrated into the Sparkoh AI platform, Arko‑T enables users to describe design intents in natural language—e.g., “generate a cup coaster” or “create a wired‑headphone storage box”—and receive a complete, parameterized CAD file that can be directly 3D‑printed or further edited. Users can iteratively refine designs by issuing additional natural‑language modifications, such as adjusting hole dimensions, without restarting the generation process.

Training data consist of roughly 1.3 million text prompt → Build123d program pairs sourced from the DeepCAD dataset, open‑source CAD projects, and internal designs. Each program is filtered through the CAD kernel to ensure it produces a non‑empty, valid entity. A crucial step is code standardization: raw CAD scripts are reorganized to expose stable patterns for features (holes, shells, arrays), extract top‑level named parameters (radius, thickness, count, spacing), and enforce a consistent construction order (sketch → extrusion → secondary features → surface treatment), making the data both executable and human‑readable.

Training proceeds in two stages. The first stage pre‑trains the model on documentation for parametric modeling tools such as Build123d, CadQuery, and OpenSCAD, teaching the language of design. The second stage fine‑tunes the model with supervised learning on the 1.3 M paired examples. This focused objective explains why a relatively modest 4 B‑parameter model can surpass much larger generalist models on the specialized task of structured CAD generation.

Arko‑T is positioned as a stepping stone toward a physical‑3D data engine. BitUnlimited plans to close the feedback loop by capturing real‑world maker interactions—design intent, geometry, assembly relationships, and post‑print modifications—and feeding them back into the model, thereby continuously improving its capabilities. The ultimate goal is to produce OpenUSD, SimReady‑level assets that can be used for robot simulation and broader Physical AI applications.

Industry analysts note that structured 3D data is a bottleneck for world‑model and embodied‑AI research. By delivering a pipeline that converts everyday maker workflows into high‑quality, editable 3D assets, Arko‑T aims to become a foundational data source for the rapidly expanding market of Physical AI, which is projected to reach trillions of dollars across consumer and enterprise segments.