ParseBench — parsers that fit on a 16 GB MacBook
The same ParseBench dimensions, restricted to parsers that run locally on a 16 GB Apple Silicon machine.
5 systems · ParseBench corpus · local runs on a 16 GB MacBook · last run August 8, 2026 · ParseBench harness · Local parsing writeup and live tracker.
Winning result
Quality — 70.1 (Chandra OCR 2)
Value — 2.0 GB (Docling)
Speed — not scored on this board.
Leaderboard
| # | System | Overall | Tables | Charts | Faithfulness | Formatting | Grounding | Footprint |
|---|
| 1 | Chandra OCR 2 Open weights | 70.1 | — | — | — | — | — | 3.5 GB |
|---|
| 2 | Qwen 3 VL 8B Open weights | 62.0 | 74.7 | 28.2 | 87.6 | 64.2 | 55.2 | 6.0 GB |
|---|
| 3 | Dots OCR 1.5 Open weights | 55.8 | 85.2 | 0.9 | 90.0 | 47.0 | 55.8 | 3.0 GB |
|---|
| 4 | Docling OSS pipeline | 50.6 | 66.4 | 52.8 | 66.9 | 1.0 | 66.1 | 2.0 GB |
|---|
| 5 | MinerU 2.5 Open weights | 45.9 | 71.7 | 1.1 | — | 4.5 | 71.5 | 3.0 GB |
|---|
- Hosted references from the same harness, for scale: LlamaParse Agentic 84.9, Gemini 3 Flash 71.0, Reducto 67.8.
- Footprint is the approximate resident memory of the 4-bit quantization or CPU pipeline we ran, not a published figure.