1. Drop in a GLB or STL. Textured as well — the colours survive the import.
2. Repair chain and report. Holes, flipped normals, degenerate triangles; the report then shows detected printing risks.
3. Map the colours onto your filaments. Onto the number actually loaded in the machine, not onto a maximum.
4. Split, pin, arrange. Dowels with clearance from the material profile, so the halves find each other again.
5. 3MF with material groups, straight to the slicer.

What a printability check checks

On the left, what a generator's check answers. On the right, what else can go wrong before the print. As of 12 August 2026, from the public documentation of a widely used generator.

Checked	Generator	Solidon3D
Watertight, holes, open edges	✓ yes	✓ yes
Non-manifold edges, degenerate faces	✓ yes	✓ yes
Volume	✓ yes	✓ yes
Wall thickness against this printer's nozzle	— no	✓ yes
Unsupported area per layer	— no	✓ yes
Islands: material with nothing below it	— no	✓ yes
Bridge spans, narrowest section	— no	✓ yes
Support needed, and the best orientation	— no	✓ yes
Does it fit your build plate	— no	✓ yes
Fits against the material profile	— no	✓ yes

Repairing alone is available free in a browser these days, and that is fine — a closed mesh is still not a part that fits. The difference sits behind the repair: dimension, fit, layer analysis, and a history in which every number can still be changed later. And what was once a hole in the mesh, Solidon3D recognises as one again — with a measured diameter, movable and resizable, without a CAD file ever having existed for it.

Your generator remains your choice

A GLB or STL file from any generator you choose goes through the same repair, sizing, printability check, splitting and pinning in Solidon3D as any other model. No CAD experience is needed.

The model behind it is TripoSG. Its source code and model card state the MIT licence; the complete licence and provenance chain of the weights and the models it embeds is currently being verified. Solidon3D sets this route up only at your request and downloads nothing unasked; external GLB and STL files remain fully usable regardless.

Local chat: which model for which card

The chat runs free of charge through Ollama on your own computer. The model has to fit entirely into graphics memory — otherwise the processor joins in and every answer takes many times as long. These four models were tested with Solidon's tools. The table shows success rates, response times and limitations:

Model	Graphics memory	Check and suite
qwen3.5:9b	7.4 GB — fully on the card from 10 GB, two to three times slower with 8 GB	Check 7 of 8, suite 21 of 39, about 6 s per request — asks back less often
gpt-oss:20b	12.9 GB — 16 GB card	Check 7 of 8, suite only 10 of 39 — for single instructions, about 7 s
qwen3:14b	13.6 GB — 16 GB card	Check 8 of 8, suite 22 of 39, about 18 s — thinks before every answer
qwen3:30b-a3b	more than 16 GB — on 16 GB the processor takes a third	Check 8 of 8, about 17 s; suite not measured

Measured on 25 September 2026 on an NVIDIA RTX 4080 with 16 GB under Ollama 0.34.3. The check is eight requests that each call for one tool; for one that is deliberately unclear, asking back is the right answer. The suite is 39 everyday tasks, from moving a part to a bracket with parts; with the whole tool set in the same window instead of the offer, qwen3:14b scored 14 there; with a window of 40 960 tokens 24, but in three times the time and no longer entirely on the card. For a local model, Solidon3D sends the intended tools with all their fields and all others in short form with each request — which makes the task a quarter as long as before. After an answer the model stays loaded for three minutes, so the next question does not wait for it to start. Not recommended, although measured: gemma4:12b (answers without end), granite4.1:8b and llama3.1:8b (guess on an unclear request instead of asking), mistral-nemo and qwen2.5-coder:14b (write the calls out as text). During setup, Solidon3D shows whether Ollama is really using the graphics card.