DeepSeek, Tencent researchers launch AI tool for CAD design

DeepSeek and researchers from Tencent, the University of Hong Kong, and Beihang University published a paper on arXiv presenting Pointer-CAD, an AI framework built on Alibaba’s Qwen 2.5 model that helps designers select edges and faces in CAD to improve 3D design accuracy.

CAD workflows start with 2D sketches, and the paper says existing AI methods either consume too many tokens or lack entity selection, which limits complex editing. Pointer-CAD combines a pointer-based representation with geometric information to let users select the most feature-consistent candidate and reduce spatial errors.

On a fillet and chamfer dataset, the authors reported an average segment error of 0.16, which they said is at least 69% lower than other methods.

The code has been open-sourced on GitHub, and reports suggest DeepSeek may also be launching its V4 model this week.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Pointer-CAD improves CAD edits with more exact entity picking

Pointer-CAD uses a pointer-based command sequence to cut confusion during CAD editing. It lets the system choose exact edges or faces in the geometry instead of vague text cues like “the left edge” 1.
The approach links the model’s output to the CAD shape, which can lower mistakes seen in methods that burn too many tokens or skip entity selection. The paper describes these gaps as a problem in complex edits 1.
The work leans into CAD-first workflows, which lines up with a collaboration that includes the University of Hong Kong (HKU) among institutions connected to related Pointer-CAD research 2.

Pointer-CAD sits within a wider move toward industry-specific foundation models

Pointer-CAD is one example of general AI techniques being tailored to specific industrial fields.
Similar “foundation model” projects include VISTA-PATH for interactive pathology segmentation (software that helps delineate tissue regions in medical images), plus earth observation models for remote sensing imagery (satellite and aerial data) 3 4.
Telecommunications research also includes Wireless Multimodal Foundation Models that jointly process RF (radio frequency) signals with other data 5.
Many industrial uses need AI that can work within hard constraints while fitting real tools and workflows, beyond general-purpose chatbots 6 7.