CADMAGEDDON agent tops frontier CAD models on a single GPU
TL;DR
- CADMAGEDDON pairs the open-weight Qwen3.8-27B model with geometric fitting tools and runs on a single 48 GB GPU under FP8 weights.
- On BenchCAD's 64³ voxel IoU, it scores 0.991 on a 1000-part subset, above vendor-reported entries for Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 Astra.
- On the MCB mechanical-parts benchmark, GMS rises from CADFather's 0.705 to 0.818 and IoU from 0.862 to 0.933 under the same evaluation.
CADMAGEDDON, an open-source mesh-to-CAD agent built on the mid-sized Qwen3.8-27B model, outperforms systems based on frontier models on all five of the paper's main benchmarks and fits on a single 48 GB GPU with FP8 weights. The preprint comes from researchers at Lomonosov Moscow State University, Innopolis University and FusionBrain Lab.
The design splits each feature in two. A vision-language model picks the structure: which plane, which profile, which operation. Geometric fitting supplies the numerical values. "The model chooses each feature's structure among named geometric entities, and fitting supplies its values," the paper reports. Algorithmic instruments detect planes, cross-sections, hole stacks, lines and arcs, give them short names, and the agent composes features by referring to those names.
On MCB, which draws on real mechanical parts, the system lifts IoU from CADFather's 0.862 to 0.933, and raises GMS, the metric that compares surface types, from 0.705 to 0.818. On CADENA-Bench the GMS jumps from 0.800 to 0.909. On the BenchCAD leaderboard, scored with the benchmark's own 64³ voxel IoU, CADMAGEDDON reaches 0.991 on a 1000-part subset. Vendor-reported entries for Claude Sonnet 5.5 (max, with tools), Claude Opus 5.5 (max, with tools) and GPT-6 Astra sit at 0.963, 0.962 and 0.959 on their 1000-part subsets, "although inputs and subsets differ," the authors note.
Task 3 of BenchCAD-2 sharpens the comparison. Frontier models on the public leaderboard do not exceed a part score of 0.59. CADMAGEDDON reaches 0.933 with CadQuery output and 0.925 with native FreeCAD programs at a reported $1.05 per part on OpenRouter pricing. The same Qwen3.8-27B wrapped in a general-purpose Qwen Code harness scores only 0.790 and returns no prediction on 2 of the 23 parts. "The instruments, not model scale, carry much of the accuracy," the authors conclude.
Limits are stated plainly. The method assumes a clean mesh of a single part; the benchmarks are all CAD-derived, so real scans and assemblies are deferred. On BenchCAD-2, every attempt leaves out the 0.7 mm walls of one part and builds at most two of the 51 through-holes of another. A reconstruction takes up to 30 or 60 model calls with long contexts, slower than single-pass learned models. Code is on GitHub. The paper lands the same day as separate Qwen3.8-27B agent work we tracked.
Originally reported by huggingface.co
Read the original article →Original headline: CADMAGEDDON Open-Source CAD Agent on Qwen3.8-27B Beats Frontier Models Across Five Benchmarks