IndustryForge-27B: A Domain-Enhanced Multimodal Foundation Model for Industrial CAD
IndustryForge-27B is a domain-enhanced multimodal foundation model built on Qwen3.5-VL-27B through unified multi-task training on 52k industrial CAD samples, which significantly outperforms both its base model and the closed-source GPT-5.4 on CAD-specific benchmarks while maintaining general capabilities to serve as a unified substrate for full-stack industrial agents.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where computers are like incredibly talented, well-read librarians who have read almost every book ever written. They can write poetry, solve math problems, and even generate code for simple video games. But there's a catch: these librarians have never stepped foot inside a factory, never held a wrench, and have certainly never looked at a complex engineering blueprint. They are great at general knowledge, but when you ask them to design a specific machine part or tell them how to operate a giant industrial robot, they start to hallucinate. They might describe a gear perfectly in words but get the size wrong, or they might try to "click" buttons on a screen that doesn't exist, because they don't understand the secret language engineers use to talk to their software.
This is the challenge of Computer-Aided Design (CAD). It's the digital art and science of building everything from tiny screws to entire airplanes using code instead of clay. For a computer to be truly useful here, it needs to do three tricky things at once: it must "see" a 2D drawing and understand the 3D shape hidden inside, it must write precise code to build that shape, and it must know how to talk to specific industrial software (like SolidWorks or Inventor) using a very old, very specific language called COM. Until now, the best general-purpose AI models were like brilliant students who failed the vocational exam—they knew the theory but couldn't do the job.
Enter IndustryForge-27B, a new kind of AI model created by researchers from Shanghai and several universities. Think of this model not as a finished robot that builds cars, but as a "super-apprentice" that has been given a massive, specialized training camp. The researchers took a powerful, general-purpose AI (called Qwen3.5-VL-27B) and fed it a carefully curated library of about 52,000 industrial examples. This library included everything from reading blueprints to writing code for single parts, assembling complex machines, and controlling industrial software. The result is a model that has learned the "secret handshake" of the engineering world.
The paper finds that this specialized training works incredibly well. When tested on four specific engineering challenges, the new model didn't just improve; it skyrocketed. It jumped from a base score of roughly 7.8% on writing code for single parts to 77.8%, and it completely crushed a top-tier, closed-source competitor (gpt-5.4) on almost every test. Perhaps most impressively, it learned to handle assemblies—putting multiple parts together—which the other models failed at almost entirely (scoring near 0%). The researchers also checked to make sure this intense training didn't make the AI "forget" how to do normal things like math or reading comprehension. It didn't; in fact, the model got slightly better at general tasks, proving that learning to be a master engineer didn't break its ability to be a smart assistant.
However, the paper is careful to set expectations. IndustryForge-27B is not a magic button that will design a new iPhone overnight. It is a foundation, a common starting point for other tools to build upon. It solves the hard problem of "can the AI even read the drawing and write the code?" so that other systems can focus on the next steps. While it is a huge leap forward, the authors admit there is still work to do, especially with complex assemblies and connecting the design process to physical simulation. But for the first time, we have a digital apprentice that actually speaks the language of the factory floor.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.