What happened when we ran three real buildings through it

For a while, every test we ran used the same bundled sample: a small house, 113 KB. It converted quickly, it shared cleanly, and it proved nothing at all.

Then we ran three real buildings through — files an architect had actually issued on live projects, 14 MB, 57 MB and 73 MB. The first one broke.

The defect

A 57 MB IFC came out the other side as a 109 MB GLB. Against a 100 MB share cap, that meant the model could be converted and viewed but could not be sent — and sending is the entire point of the product.

Nothing had flagged this, because the only IFC anyone had ever converted was the 113 KB sample. It sat comfortably inside every limit, so every limit looked fine.

Worth saying plainly: an export getting bigger than its input is not unusual. IFC describes a wall parametrically; a GLB has to describe the triangles. If a model repeats the same window a few hundred times, a naive conversion writes those triangles a few hundred times.

What actually fixed it

Two things, both about repetition rather than resolution:

Instancing. Recognise that those few hundred windows are the same geometry in different places, store it once, and reference it. This is where most of the reduction comes from on real buildings, because real buildings are extremely repetitive.

Meshopt compression on what is left.

Neither degrades what you see. This is not decimation — we are not throwing away detail and hoping nobody measures anything.

The numbers

| Real IFC in | GLB out | Reduction | |---|---|---| | 14 MB office | 1.94 MB | 13× | | 57 MB school | 6.79 MB | 16× | | 73 MB architectural | 14.6 MB | 9× |

The 73 MB file is the one that matters most. That is a large architectural model, and 14.6 MB is a size a client can open on mobile data without thinking about it.

Why we publish this

Because "it handles large models" is a claim, and 73 MB → 14.6 MB is a measurement. The second one you can check.

It is also a reminder we keep having to take seriously ourselves: testing on the sample proves the sample works. The bundled house had passed every test for months while the product could not have survived its first real project. If you are evaluating this — or any viewer — do it on one of your own files, from a live job, at the size you actually work at.

That offer stands, by the way. Send us an IFC or STEP from something current and we will open it and send back a link you can try on site.