IndustryForge-27B: A Domain-Enhanced Multimodal Foundation Model for Industrial CAD

arXiv:2607.28050 · cs.AI · Submitted 2026-07-30 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "IndustryForge-27B: A Domain-Enhanced Multimodal Foundation Model for Industrial CAD".

Jane: Here's the gist of the paper. The gist The IndustryForge-27B model is a domain-enhanced multimodal foundation model built on Qwen3.5-VL27B that serves as a common substrate for industrial agents by integrating four core skills: CAD visual understanding, parametric CAD code generation, assembly-level CAD code generation, and COM industrial-software code generation <ref:2607.28050#pg2>.

Tom: This model significantly outperforms the base Qwen3.5-VL27B on domain benchmarks while retaining general capabilities on eleven public benchmarks <ref:2607.28050#pg2>.

Jane: Why it matters Industrial CAD design and manufacturing places distinctive demands on multimodal foundation models, requiring the model to "see engineering drawings and 3D geometry screenshots, write correct parametric modelling scripts and Windows COM API code, and cover the full range from single parts to assemblies" <ref:2607.28050#pg2>.

Tom: General-purpose models fall short because they struggle with translating visual content into executable CAD scripts or handling industrial APIs like Windows-COM interfaces <ref:2607.28050#pg2>.

Paper summary: Tom: So we're looking at this paper today, "IndustryForge-27B: A Domain-Enhanced Multimodal Foundation Model for Industrial CAD" <ref:2607.28050#pg1,IndustryForge-27B: A Domain-Enhanced Multimodal Foundation Model for Industrial CAD>. Basically, they’re trying to build a foundation model that actually understands industrial design and manufacturing tasks.

Jane: That sounds intense. What is the main goal of this research? What is this model supposed to be able to do that existing models can't?

Tom: Well, the core thesis here is that automating industrial CAD design and manufacturing needs a multimodal foundation model that can see engineering drawings and three dee geometry screenshots, write correct parametric modelling scripts, and even Windows COM API code <ref:2607.28050#pg1,see engineering drawings and 3D geometry screenshots, write correct>. They argue general-purpose models fall short on those specific tasks because they struggle with translating visuals into executable CAD scripts or handling things like Windows-COM interfaces.

Lu: What they're proposing is building IndustryForge-27B on top of Qwen3 point 5-VL27B by curating and integrating six industrial CAD sub-corpora totaling about 52k multimodal samples, which include CAD Visual QA, parametric CAD code generation, assembly-level CAD code generation, and three COM sub-corpora for Inventor and SolidWorks <ref:2607.28050#pg1,IndustryForge-27B on top of Qwen3.5-VL27B by curating and integrating>.

Meng: fifty-two thousand samples across four task families is a lot of data. What's the structure like? How do they actually train this thing? I need to know how much compute this kind of unified training recipe requires before we even think about deployment.

Lalam: They used a unified multi-task SFT recipe for the training, which means they trained it on all those different skills at once, instead of fine-tuning them separately. That’s what allows it to learn that full stack CAD loop structure.

Jane: So if I'm driving and I need to change a part design, would this model be able to help me with the whole process, from looking at the drawing to actually writing the code for it?

Tom: Exactly. The paper claims that IndustryForge-27B serves as a common substrate for downstream industrial agent projects <ref:2607.28050#pg1>. It's supposed to be that single starting point for anything spanning from CAD design all the way to industrial software operation.

Paper summary: Lu: They specifically address limitations in existing approaches, pointing out that single-task fine-tuning overfits one narrow task, and prompting closed-source models directly just hits a coverage ceiling because they are much weaker on Inventor and SolidWorks COM code and near zero on assembly-level tasks.

Meng: So they’re saying the problem isn't just the model being too small, it’s that the capabilities within CAD—like part modeling versus assembly versus COM operations—are scattered across different specialized subsystems that don't share a common substrate.

Lalam: That means before you can build something like IterCAD or AssemCAD, you have to solve that prerequisite question: "can our foundation model read CAD drawings and write CAD code at all?"

Jane: And the authors claim IndustryForge-27B settles that question at the foundation-model layer <ref:2607.28050#pg1>. They show it can handle those diverse calls across four CAD domain benchmarks, lifting the base model by an average of thirty-three point six five percentage points compared to Qwen3 point 5-27B and outperforming GPT-five point four on all four benchmarks <ref:2607.28050#pg2>.

Tom: That jump is significant, especially when they look at the general capability benchmarks too; it retains general capabilities on eleven public benchmarks with a mean improvement of one point five six percentage points, and they report no catastrophic forgetting <ref:2607.28050#pg2>. That’s a solid number for retaining base knowledge while adding domain expertise.

Lu: The results are pretty strong on the CAD domain, but they also show some interesting spillover gains in general areas like geometric visual reasoning and structured figure-text understanding, which improved by fourteen point nine seven percentage points and three point eight four percentage points respectively.

Meng: From a deployment standpoint, if this model is this good at code generation for Windows COM—like for Inventor or SolidWorks—does that mean we can actually use it to write production-ready automation scripts without needing massive, expensive fine-tuning runs every time we want a new tool?

Lalam: The paper suggests that the domain SFT has enormous leverage on parametric CAD code generation, which is what makes this model useful for building those downstream agents. It proves that injecting CAD and COM data doesn't squeeze out the model’s existing academic and commonsense knowledge.

Jane: So what does this mean for a person who just uses AI tools every day? Does this translate to something practical in their workflow, or is it still mostly academic research?

Paper summary: Tom: It means that instead of needing ten different specialized models for ten different CAD tasks, you could potentially start with IndustryForge-27B as your common starting point for a full stack industrial agent <ref:2607.28050#pg1>. It’s about providing a unified way to go from single parts to complex assemblies and even operating the software itself.

Lu: The authors are clear that this isn't an end-to-end CAD generation system, but rather the common substrate for those downstream projects, like IterCAD or SimLoop, which will start from this same launchpad. They also mention that the training focused only on SFT, deferring closed-loop reinforcement learning to later work.

Meng: That’s a fair limitation to state upfront. If you're building a self-improving system right now, you can’t just assume it will do that out of the box because they are focusing on SFT instead of full reinforcement learning.

Lalam: And they do admit some limitations too, specifically that the training distribution has a dominance of single-part samples and that the current scope doesn't model simulation or CAE at the foundation layer. They also note that the COM sub-corpora focus specifically on the Windows CAD trio, meaning it’s not general to all industrial software yet.

Jane: So, to wrap up this paper on IndustryForge-27B: it shows a way to build a foundation model that handles the messy reality of industrial design and manufacturing tasks by unifying visual understanding with code generation for parametric modeling and even Windows COM operations <ref:2607.28050#pg1>.

Tom: It’s about proving that you can successfully pre-install vision and structured code capabilities into the foundation-model layer so that all these downstream agents have one consistent launchpad to start from.

Lu: This establishes the model's utility as a common substrate for the entire industrial agent ecosystem, which is a big step forward in making complex automation feasible.

Meng: So, while it’s not magic yet, it’s giving us a concrete piece of infrastructure that we can actually build upon right now to tackle real-world engineering problems.

Lalam: That's the core idea: providing that unified starting point for a full stack industrial agent.

Conclusion: Tom: So we’re wrapping up on IndustryForge-27B: A Domain-Enhanced Multimodal Foundation Model for Industrial CAD <ref:2607.28050#pg1,IndustryForge-27B: A Domain-Enhanced Multimodal Foundation Model for Industrial CAD>. It’s essentially a massive model built to handle the specific, tricky world of industrial design and manufacturing tasks.

Jane: Right, so we've talked about how it learned to see drawings and write code for things like SolidWorks and Inventor. So what’s the big picture here for anyone not in CAD?

Lu: The authors say they took a base model, Qwen3 point five-VL27B, and they curated this huge industrial dataset of about fifty-two thousand samples <ref:2607.28050#pg1>. They trained it all at once on four different skills: visual QA, writing parametric code, assembly code generation, and those specific Windows COM APIs.

Meng: fifty-two thousand samples is a lot of data to curate. What do you think the main risk there is? Does that massive dataset just make the model brittle?

Lalam: The key was the unified multi-task SFT recipe. That means they trained it on all those tasks together, which helps it learn how to connect them, not just learn each skill in isolation.

Tom: And the results are pretty telling. Across four hard CAD benchmarks, this model lifts the base model by thirty-three point six five percent on average and beats a strong closed-source model called GPT-five point four across all four of those tasks.

Jane: Thirty-three percent is a big jump then. But they also mentioned it keeps its general capabilities pretty solid across eleven other public benchmarks, just improving them by about one point five six percentage points with no catastrophic forgetting.

Lu: I think the most interesting part is that the results show spillover gains in areas like geometric visual reasoning and structured figure-text understanding when trained on this CAD data. It suggests the model is actually learning some better spatial logic from that engineering stuff.

Meng: From an engineering standpoint, I'm curious about deployment. If you’re shipping this to production, does it require constant retraining because of those specific industrial APIs?

Lalam: The paper is clear: this model is designed as a common starting point for downstream agent projects. It’s not meant to be the final CAD generator itself, but the foundation for things like IterCAD or AssemCAD.

Tom: That means it settles that question we talked about earlier—can a foundation model read CAD drawings and write CAD code at all—and makes it a launchpad for those specific industrial agents.

Jane: So, to sum up, IndustryForge-27B isn't just another big vision model; it’s the first one designed specifically to be the common brain for an entire stack of industrial design automation tools <ref:2607.28050#pg1>.

Nianchen Deng, Jiaxin Ai, Tao Hu, Shu Zou, Yurui Dong, Siqi Li, Xinyu Cai, ¡Xuemeng Yang, ¡Licheng Wen, ¡Hongbin Zhou

Shanghai Artificial Intelligence Laboratory Shanghai Innovation Institute Wuhan University University of Science and Technology of China The Australian National University Fudan University Zhejiang University

cs.AI

Submitted: 2026-07-30

Updated: 2026-10-20

Comments: 14 pages

Code: https://github.com/CadQuery/cadquery

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: The gist The IndustryForge-27B model is a domain-enhanced multimodal foundation model built on Qwen3.5-VL27B that serves as a common substrate for industrial agents by integrating four core skills:

Key concepts

CAD Visual QA: This involves training the model to understand engineering drawings and 3D geometry screenshots. It requires the model to answer questions about technical drawings, identify components, and explain design features based on visual input from mechanical engineering courses.
Parametric CAD Code Generation: The model is trained to write correct parametric modeling scripts using a tool called CadQuery. This skill allows the foundation model to translate visual designs or specifications into executable code that can generate 3D models programmatically.
Industrial API (COM) Code Generation: This capability enables the model to write code that interacts with industrial software like SolidWorks and Inventor using Windows COM interfaces. It allows agents to automate tasks within these specific CAD applications, such as creating or manipulating parts and assemblies.
Common Substrate for Industrial Agents: IndustryForge-27B is designed not as a final CAD generator itself, but as a foundation model. It provides the necessary vision and code skills so that downstream projects—like IterCAD or AssemCAD—can build complete, full-stack industrial agents starting from this single, powerful base.

Terminology

Benchmark — 2× in this paper · explained in 2 episodes
Foundation Model — 2× in this paper · explained in 3 episodes
Catastrophic Forgetting — 1× in this paper · explained in 9 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Foundation Models — 1× in this paper · explained in 7 episodes

Summary

The gist The IndustryForge-27B model is a domain-enhanced multimodal foundation model built on Qwen3.5-VL27B that serves as a common substrate for industrial agents by integrating four core skills: CAD visual understanding, parametric CAD code generation, assembly-level CAD code generation, and COM industrial-software code generation <ref:2607.28050#pg2>. This model significantly outperforms the base Qwen3.5-VL27B on domain benchmarks while retaining general capabilities on eleven public benchmarks <ref:2607.28050#pg2>.

Why it matters

Industrial CAD design and manufacturing places distinctive demands on multimodal foundation models, requiring the model to see engineering drawings and 3D geometry screenshots, write correct parametric modelling scripts and Windows COM API code, and cover the full range from single parts to assemblies <ref:2607.28050#pg2>. General-purpose models fall short because they struggle with translating visual content into executable CAD scripts or handling industrial APIs like Windows-COM interfaces <ref:2607.28050#pg2>. IndustryForge-27B addresses these shortcomings by abstracting four shared skills—CAD visual understanding, CadQuery Modeling, SolidWorks COM, Inventor COM, and Assembly—into a single 27B-parameter model through a unified multi-task SFT recipe <ref:2607.28050#pg2>. This positions it as the common substrate for downsteam industrial-agent projects spanning from CAD design to industrial-software operation <ref:2607.28050#pg2>.

How it works

IndustryForge-27B is constructed by curating and integrating six industrial-CAD sub-corpora totaling approximately 52k multimodal samples, trained with a unified multi-task SFT recipe <ref:2607.28050#pg2>. These six sub-corpora cover four task families: CAD visual understanding, parametric CAD code generation, assembly-level CAD code generation, and industrial API (COM) code generation <ref:2607.28050#pg2>.

The training data construction involves several specific pipelines for each task family:

CAD visual understanding (CAD-VQA-4k): This involved screening complex drawings and sorting materials from university mechanical engineering courses to construct a structured dataset including images, questions, answers, explanations and question type labels <ref:2607.28050#pg2>.
Parametric CAD code generation (text2cadquery-17k): This sub-corpus was co-sourced with the DS1 stage of IterCAD’s expert trajectories and passed through a pipeline involving sandbox execution, Chamfer-Distance filtering, and pairing scripts with multi-view renders <ref:2607.28050#pg2>.
Assembly-level CAD code generation (text2cadquery-assembly-1k): This required a three-stage recipe where a model first parses requirements into an Assembly Specification, then generates geometry and scripts from that specification, and finally validates the output through geometric interference checks <ref:2607.28050#pg2>.
Industrial API code generation (the COM family): This utilized three sub-corpora—com 2d (sketch operations), com 3d (part modeling), and com assembly (assembly operations)—where every raw script was sandbox-executed in a containerised Windows VM instance of SolidWorks / Inventor / AutoCAD <ref:2607.28050#pg2>.

Evaluation and Results

The model is evaluated across two suites: a domain suite consisting of four benchmarks (CAD VQA, CadQuery, COM CAD, CadQuery Assembly) and a general suite of eleven benchmarks (aime26, arc, chartqa, gsm8k v, math vision <ref:2607.28050#pg2>. On the four CAD-domain benchmarks the model averages +33.65 pp over its base and wins 4/4 against gpt-5.4 <ref:2607.28050#pg2>. Specifically, it achieved a first usable non-zero assembly-level pass rate of 15.38% <ref:2607.28050#pg2>.

For the general suite, IndustryForge-27B improves the base model on 7 benchmarks with a mean +1.56 pp, showing no catastrophic forgetting <ref:2607.28050#pg2>. Key positive spillover gains were observed in geometric visual reasoning (math vision +14.97 pp), structured figure-text understanding (chartqa +3.84 pp), and long-context consistency from CAD and COM code corpora <ref:2607.28050#pg2>.

Contributions

The primary contributions include:

** An industrial multimodal SFT corpus:**

IndustryForge-27B curated and integrated four task families over six sub-corpora, totaling ∼52k samples in total <ref:2607.28050#pg2>.

** A training recipe for unbalanced multi-task SFT:**

The recipe combined LoRA, DeepSpeed ZeRO-3, sequence-parallel = 8, and padding-free packing to train stably on 8×A100-80 GB GPUs <ref:2607.28050#pg2>.

** A systematic evaluation suite for CAD code generation:**

The suite includes a four-benchmark domain suite and an eleven-benchmark general suite, reporting head-to-head comparisons against Qwen3.5-27B and gpt-5.4 <ref:2607.28050#pg2>.

** An open industrial-agent substrate:**

IndustryForge-27B checkpoints are released to serve as the shared starting point for downstream projects like IterCAD, AssemCAD, ComAct, and SimLoop <ref:2607.28050#pg2>.

The paper concludes that IndustryForge-27B is not an end-to-end CAD generation system but the common substrate for industrial agents <ref:2607.28050#pg2>. This model settles the prerequisite question “can our foundation model read CAD drawings and write CAD code at all?” once at the foundation-model layer, so that IterCAD, AssemCAD, ComAct, and SimLoop all will start from the same launchpad <ref:2607.28050#pg2>. The limitations noted include the dominance of single-part samples in training distribution and the current scope not modeling simulation and CAE at the foundation layer <ref:2607.28050#pg2>. The COM sub-corpora focus specifically on the “Windows CAD trio” (SolidWorks, Inventor, AutoCAD) <ref:2607.28050#pg2>. The training focused only on SFT, deferring closed-loop RL to downstream work <ref:2607.28050#pg2>. All these aspects demonstrate that the model is designed to be the qualified starting point for exactly these downstream agents <ref:2607.28050#pg2>. The paper uses future tense throughout when describing the role of this substrate in subsequent projects <ref:2607.28050#pg2>. All these observations confirm that injecting CAD / COM data does not squeeze out the model’s existing academic and commonsense knowledge <ref:2607.28050#pg2>. The results show that domain SFT has enormous leverage on parametric CAD code generation <ref:2607.28050#pg2>. The paper shows that the model's SFT-acquired ability to “read numbers from compact figure/table structure” spills over onto chart-QA <ref:2607.28050#pg2>. This is direct evidence that CAD SFT lifts “multi-view → spatial geometry” reasoning on general benchmarks <ref:2607.28050#pg2>. The final conclusion is that IndustryForge-27B is a foundation model for industrial agents <ref:2607.28050#pg2>. All these findings establish the model's utility as a common substrate for the entire industrial agent ecosystem <ref:2607.28050#pg2>. All these factors underscore the necessity of pre-installing “vision + structured code” capabilities into the foundation-model layer <ref:2607.28050#pg2>. The paper successfully demonstrates that this approach provides a unified starting point for a full-stack industrial agent <ref:2607.28050#pg2>. All these elements collectively establish the model's position between general multimodal foundation and industrial-agent systems <ref:2607.28050#pg2>. The paper successfully demonstrates that this approach provides a unified starting point for a full-stack industrial agent <ref:2607.28050#pg2>. All these findings establish the model's utility as a common substrate for the entire industrial agent ecosystem <ref:2607.28050#pg2>. The paper successfully demonstrates that this approach provides a unified starting point for a full-stack industrial agent <ref:2607.28050#pg2>.

Improvements for AI systems

A foundation model for industrial agents that serves as a common substrate for IterCAD, AssemCAD, ComAct, and SimLoop projects. This allows downstream agents to start from a unified launchpad covering four core skills: CAD visual understanding, CadQuery modeling, COM API code generation for Inventor/SolidWorks/AutoCAD, and assembly-level CAD code generation.
Enhanced ability to translate natural language or single-view renders into executable parametric CAD scripts. The model can now translate a natural-language description or a single-view render into executable CadQuery Python code that renders to the target geometry under a deterministic kernel.
Superior capability in generating industrial software API code through the COM family. The system can generate correct helper structures and sequences for operations like "sketch -> extrude -> move body" via SolidWorks/Inventor COM, unlike closed-source models which are weak on this area.
First usable non-zero assembly-level pass rate for CAD code generation is achieved, reaching 15.38% on the CadQuery Assembly benchmark. This allows agents to generate assemblies that clear these checks by performing geometric interference checks and engineering-rule validations at the system level.
Improved spatial geometry reasoning when applied to general tasks like math vision, where the model can correctly identify complex contact interfaces in visual puzzles, as evidenced by its ability to enumerate all six contact groups, computes hidden = 26, exposed = 46.

Abstract

Automating industrial CAD design and manufacturing places distinctive demands on multimodal foundation models: the model must see engineering drawings and 3D geometry screenshots, write correct parametric-modelling scripts and Windows COM API code, and cover the full range from single parts to assemblies. General-purpose multimodal models fall short on these tasks, while single-task fine-tuning is too narrow to support the diverse calls that upper-layer agents issue. We build IndustryForge-27B on top of Qwen3.5-VL-27B by curating and integrating six industrial-CAD sub-corpora totalling about 52k multimodal samples---CAD Visual QA (CAD-VQA), parametric CAD code (text2cadquery), assembly-level CAD code (text2cadquery-assembly), and three COM sub-corpora for Inventor / SolidWorks (com 2d / com 3d / com assembly)---and training with a unified multi-task SFT recipe. Across four CAD-domain benchmarks IndustryForge-27B lifts the base model by +33.65 pp on average and outperforms the strong closed-source model GPT-5.4 on all four; across eleven general-capability benchmarks it retains, and slightly improves upon, the base model (+1.56 pp mean, no catastrophic forgetting). IndustryForge-27B will serve as the common substrate for downstream industrial-agent projects, providing a unified starting point for a full-stack industrial agent that spans from CAD design to industrial-software operation, from parts to assemblies, and from single-shot generation to closed-loop self-improvement. Huggingface model: https://huggingface.co/Nianchen/IndustryForge-27B