Which model fits your machine?
Three questions set the weights. Eight axes score every parser. Your memory budget decides what runs here. One pick, with the reasons.
Where can your documents go?
- Never off this device — On-device models only.
- Our own servers — On-device or self-hosted.
- A vendor with no-training terms — Hosted is fine when the vendor does not train on your files.
- Anywhere — Quality first. Any hosted parser.
What do your documents contain?
- Tables — Financials, schedules, line items.
- Charts and figures — Bar, line, and pie charts as data.
- Scans and forms — No text layer. Boxes must land on the page.
- Long text with headings — Reports, filings, contracts.
When they conflict, which wins?
- Quality — Take the slower parser.
- Balanced — Weigh speed like any quality axis.
- Speed and cost — Fast parsers under 1¢ per page.
Eight axes
- Tables — Table structure: cells, merged cells, and headers survive the parse.
- Charts — Charts and figures come back as data, not as a caption.
- Faithfulness — Text comes back complete. No dropped rows, swapped digits, or invented values.
- Formatting — Headings, lists, and reading order survive the parse.
- Grounding — Layout boxes land on the right region of the page.
- Privacy — Where the document bytes go. Derived from dataHandling.
- Runs here — Fit on the user's own machine. On-device models score by memory headroom. Self-hosted containers get partial credit. Hosted APIs score zero.
- Speed — Single-stream pages per minute: one request or one machine at a time. Hosted APIs can run many streams in parallel; this axis does not credit that.
Quality scores are recited from the ParseBench leaderboard snapshot (arXiv 2604.08538). Privacy, runs here and speed come from the catalog facts and the reader's machine.
See every benchmark · Download for Mac
