Measured · 2026-10-29
We measured every 3D model we sell
470 published models, 14.9 GB of GLB. Not estimates, not vendor specs — we read the header and accessor tables of every file and counted. Here is what a 3D model actually weighs in 2026, and where that weight goes.
The distribution, not the average
Averages hide everything interesting here. The middle half of the library sits between 15.9 MB and 23.3 MB, but the tail runs to 98.2 MB — and triangle counts spread over six figures.
| Metric | Min | P25 | Median | P75 | Max |
|---|---|---|---|---|---|
| File size (MB) | 11.2 | 15.9 | 18 | 23.3 | 98.2 |
| Triangles | 248,886 | 474,196 | 605,485 | 839,073 | 1,500,288 |
| Bytes / triangle | — | 31 | 36.5 | 49.2 | — |
115 of 470 files carry more than a million triangles. At 0.2 mm layers on a 100 mm model that is over 2,000 triangles per printed layer.
The size spread is encoding, not detail
One catalog, two encodings. The compressed half is 4.7× smaller in total and 1.7× smaller per triangle — while carrying fewer triangles, so none of that saving is paid for in detail. The rest of the gap is simply less geometry, and the per-triangle part is purely how the buffers are written to disk.
Positions stored as 16-bit integers with a node-level scale; buffers de-interleaved and entropy-coded.
Written by glTF-Transform v4.5.0
Full float32 positions and normals, one buffer per attribute, no compression extension.
Written by Khronos glTF Blender I/O v4.0.43
What this means if you are buying a print
Triangle count is not print resolution. A 100 mm model at 0.2 mm layers is about 500 layers, and the median file here averages ~1,211 triangles per layer. A slicer reduces each layer to a polyline with tens to low hundreds of points, so the finer half of that geometry never reaches the nozzle. Going from 600K to 1.5M triangles mostly buys slicing time, not surface.
Bigger files are not safer files. The most common way a print fails is a wall too thin or a mesh that is not closed — neither is visible in a file size. That is why we quote printability from a review, and why this audit does not claim to know it: our measurement reads geometry tables, not solid bodies.
Scale is the expensive mistake. glTF stores no unit at all — it is a file format convention, not a guarantee. A model authored in metres and opened as millimetres arrives 1000× too large, and nothing in the file size warns you. It is the one failure that wastes a whole build plate, which is why every quote here is pinned to a measured bounding box rather than a declared one.
How this was measured
- 1.Sample: every published, non-seed model in the ModelToMake catalog that has a GLB file — 470 of 470. No sampling, no exclusions.
- 2.Measurement: for each file we HTTP range-read the 32-byte GLB header, then the JSON chunk (metres, mesh, accessor tables) and never the binary buffer. Triangle counts are summed from index accessor counts; vertex counts from POSITION accessor counts.
- 3.What this method cannot see: watertightness, wall thickness, overhangs and true physical dimensions. Those need the full binary buffer and a solid-body analysis, which is why we do not quote them here.
- 4.Tool: scripts/glb_benchmark.py in our public repository. The per-model CSV behind every figure is downloadable above.
Per-model dataset (470 rows)
model id, file size, triangles, vertices, encoding, generator · CC BY 4.0
What we are not claiming
- No cross-engine benchmark. Every file in this catalog comes from one generation pipeline. We have no second engine to compare against, so we publish no render-time or success-rate comparison — a benchmark needs two real subjects.
- No printability verdict. Watertightness, wall thickness and overhang angles require solid-body analysis of the full buffer. Counting accessors cannot see them, so we do not quote them.
- No industry claim. This describes our own 470 models. It is not a survey of the market.
Questions this page answers
- How big is a 3D model file in 2026?
- In our library the median published model is 18 MB, with a middle half between 15.9 MB and 23.3 MB. The smallest is 11.2 MB and the largest 98.2 MB. File size is driven by texture count and how the mesh buffers were encoded, not by how detailed the object looks.
- How many triangles does a 3D-printable model need?
- Our median file carries 605,485 triangles. At a 0.2 mm layer height a 100 mm tall model is about 500 layers, so that averages roughly 1,211 triangles per layer — while a slicer turns each layer into a closed polyline with tens to low hundreds of points. Most of the geometry is resolved away before the nozzle sees it, so file weight and print resolution are close to unrelated.
- Does a bigger 3D file mean a better print?
- No. Print quality is set by layer height, nozzle diameter and material, not by triangle count. In this audit the smaller half of the library carried fewer triangles per file and still prints at the same resolution — the difference is buffer encoding (quantized positions plus meshopt compression), which is invisible in the finished part.
- Why do you publish this data?
- Because most 3D-printing advice about file sizes is guesswork. These are counts taken directly from the bytes of 470 files with a script you can re-run, and the per-model table is downloadable. If a number here looks wrong to you, check it against the CSV.