Voxen: Qwen3-8B Assembly-Aware CAD on AMD MI300X

The Problem: AI-to-3D tools output single meshes—visually interesting but useless for real CAD workflows. Engineers can't edit parts, assign materials, or export to Fusion 360 or SolidWorks. Our Solution: Voxen generates assembly-aware models where every component is a discrete, labeled part with dimensions, materials, and design intent. Built on AMD MI300X with fine-tuned Qwen3-8B for structured CAD generation. How It Works: 1.User describes assembly in natural language 2.Fine-tuned Qwen3-8B generates Zod-validated JSON on AMD MI300X 3.React Three Fiber renders each part as independent 3D mesh 4.User inspects interactively—selected part solid, others wireframe 5.Export parts or full assembly as STL/OBJ/STEP Fine-Tuning: We fine-tuned Qwen3-8B-Instruct using PyTorch on AMD MI300X with ROCm. Created synthetic dataset of 100+ mechanical assemblies (robotic arms, grippers, hinges) with ground-truth JSON schemas. Used LoRA for parameter-efficient training, reducing time by 60% while maintaining accuracy. Tech Stack: AMD MI300X for training/inference • PyTorch + ROCm pipeline • Qwen3-8B-Instruct (fine-tuned) • vLLM inference • Next.js 16 + TypeScript • React Three Fiber • Zod validation • Flask + Pydantic backend Key Innovation: Unlike blob generators, Voxen outputs CAD-ready assemblies. Fine-tuning on domain data improved JSON accuracy by 40% and reduced invalid outputs by 85%. Each part is independently editable and compatible with manufacturing workflows. AMD MI300X Impact: MI300X's 192GB HBM3 enables full-precision fine-tuning without quantization. 5.3 TB/s bandwidth generates 6+ part assemblies in under 8 seconds—critical for iterative design. ROCm's mature PyTorch support made training seamless.