Vision models in context
The goal is to turn a substation single-line diagram into data an engineer can trust: every symbol, tag and voltage, and which ends meet at each connection, in the IEC 61850 SCL data format. This page puts the models tested here next to public rankings of vision models.
What this task asks of a model
- Read small text exactly: tags like CB-735-11 and voltages, copied as drawn.
- Tell similar symbols apart: a breaker from a disconnector from an earthing switch.
- Trace the wiring: which equipment connects to which.
- Answer completely, every run: a reply that runs out of room gives no result.
General leaderboards mostly measure photos, charts, documents and counting. None of them measures wiring. That is why the ranking below combines them with one engineering-diagram benchmark and with the runs on this site.
Ranking for drawing → structured data
Public figures as published on the dates in the sources below, not re-measured here. "This site" is computed from the saved runs: one synthetic drawing in two renderings, scored against one answer key.
| # | Model (maker) | Weights | Public evidence | This site |
|---|
Open-weight models can be downloaded and run privately, which matters for drawings that may not leave a company. "Open" here does not mean MIT: each has its own license.
What an engineering-diagram benchmark found
Enginuity (2026) tests vision models on 2,056 engineering diagrams paired with parts tables, with exact ground truth.
- Wiring diagrams were the hardest type: recall 0.620, because they have no numbered callouts to anchor each part.
- Models find parts better than they describe them: recall reached 0.865 (Claude Opus 4.7 with reasoning, n=494 diagrams), but exact-wording scores stayed between 0.034 and 0.179.
- Structure needs reasoning: GPT-5.2 often returned malformed JSON for long parts lists without step-by-step reasoning; with it, recall rose from 0.608 to 0.817 (n=494).
- Question answering is unsettled: on 60 diagram questions, the top models' scores overlapped too much to rank them reliably.
What it means for this evaluation
The next step is harder and real drawings. The shortlist is the top of the table above, plus the open-weight models that can run privately.
Sources
- Enginuity: a dataset and benchmark for vision-language understanding of engineering diagrams (arXiv 2606.03410, 2026)
- Arena leaderboard, October 2026 (BenchLM, snapshot of 23 Oct 2026; human-vote Elo across all tasks, not vision only)
- Roboflow Vision Evals (five tasks, 65 models, updated 26 Oct 2026)