Meshy announced on 13 November 2026 a company-funded benchmark of AI-generated 3D models physically tested at the University of Massachusetts Amherst’s ADDFab facility. The company reported fewer major missing features after its Make Printable preparation, while disclosing that the selected FDM challenge set does not establish performance across all prompts. Meshy’s announcement explicitly says facility participation is not product endorsement.
The November 13 release publicised results already described in a Meshy blog post dated November 10. It is an announcement of study results; it does not establish that the printing took place on either publication date.
What the printability test found
Meshy reports that 11 of 25 prints in its Meshy 7 comparison workflow had major missing features or holes, compared with two of 25 after Make Printable. That corresponds to 44% and 8%, respectively. The widely promoted 92% figure describes the 23 prepared prints without those major completeness defects, rather than prints without any defect.
The company also identifies two prepared models that still had major missing features or holes: a six-legged insect and a long-finned tropical fish. The physical test used fused deposition modelling (FDM); the release discusses other printing processes as possible future testing.
Who funded the study and performed the testing
According to Meshy’s study overview and collaboration disclosure, Meshy commissioned and funded the work, supplied six Bambu Lab A1 printers and consumables, generated the models, prepared digital geometry and performed the software analysis.
ADDFab handled file preparation and slicing, printing, support removal and objective grading. The lab developed the grading rubric and recruited volunteers for the preference evaluation. Engine identities were hidden during file preparation, printing and grading. The physical set comprised 25 prompts across five workflows, producing 125 specimens under matched printer settings and filament. The disclosure distinguishes that laboratory role from endorsement or an audit of Meshy’s software.
Why the result has a narrow scope
The full whitepaper clarifies that the physical Meshy 7 baseline used standard repair. It was not the raw Meshy baseline used in the separate software assessment. Its severe-completeness flag applies when any grader assigns the highest completeness-defect score; it does not mean a printer aborted.
The document describes deliberate challenge selection, prioritising prompts predicted to produce overly thin competitor models. It also records exclusions after preparation damaged models and two seed choices made by visual inspection. Those choices create a selection-bias concern when interpreting the headline rate. There was one physical specimen per prompt and workflow; three graders were not three repeated prints.
The supplied paired analysis did not statistically distinguish Make Printable from Rodin on total defect scores. The downloadable document is marked “Internal review draft” and revised October 29, 2026. These details constrain the findings’ generalizability beyond the selected test.
What this means for university research facilities
For university labs considering AI-generated models, the useful development is a disclosed physical evaluation alongside software geometry checks. Our interpretation is that local evaluation should preserve that distinction: model geometry, observed print defects and user preferences answer different questions. For methods reproducibility, retain the preparation workflow and fabrication conditions when comparing results. Research offices describing the collaboration should also preserve the sponsor disclosure and the explicit boundary around the university’s role.
Sources
- Meshy.ai: November 13 benchmark announcement, PR Newswire. Checked 2026-11-15.
- Meshy: Same Prompt, More Complete Prints, November 10 blog post. Checked 2026-11-15.
- Meshy: Printability study results and collaboration disclosure. Checked 2026-11-15.
- Meshy: Printability of AI-generated 3D models, revised October 29 whitepaper. Checked 2026-11-15.
