Which resume extraction model handles handwritten resumes best?
9 of the 300 resumes are handwritten resumes. The typical system scores 31.7 points lower on them. Best: LlamaParse Extract Agentic Plus, 67.8 F1.
Resumes written by hand and scanned or photographed. There is no text layer: every word has to be read from the handwriting. Best document API: LlamaParse Extract Agentic Plus (67.8). Best vision LLM: Claude Opus 4.8 (54.7). GLM 5.3 Flash ranks 14 places higher here than overall.
Every system on handwritten resumes
F1 on the 9 handwritten resumes, against each system’s F1 on all 300 resumes. Cost and time are for all resumes.
| Model | F1 here | F1 overall | Gap | Per 1,000 pages | Median time |
|---|---|---|---|---|---|
| LlamaParse Extract Agentic Plus | 67.8 | 87.2 | −19.3 | $54 | 70 s |
| LlamaParse Extract Agentic | 63.5 | 85.5 | −22.1 | $23 | 55 s |
| Claude Opus 4.8 | 54.7 | 85.5 | −30.8 | $90 | 31 s |
| GLM 5.3 Flash | 54.6 | 59.2 | −4.6 | $0.80 | 46 s |
| Qwen 3.8 Flash | 53.5 | 84.1 | −30.6 | $2.98 | 122 s |
| GPT-5.6 | 53.2 | 84.7 | −31.5 | $74 | 38 s |
| Qwen 3.5 35B | 52.7 | 77.9 | −25.2 | $3.91 | 78 s |
| Extend Extract | 52.5 | 81.0 | −28.5 | $12 | 31 s |
| Qwen 3.8 27B | 51.6 | 83.8 | −32.2 | $9.59 | 36 s |
| GPT-5.4 Nano | 51.3 | 61.9 | −10.5 | $2.76 | 20 s |
| GPT-5.5 | 51.0 | 83.5 | −32.5 | $87 | 48 s |
| Qwen 3.5 9B | 50.6 | 76.0 | −25.4 | $1.31 | 138 s |
| Gemini 3.5 Flash | 50.3 | 85.5 | −35.2 | $35 | 43 s |
| Gemma 4 26B | 50.3 | 62.8 | −12.5 | $0.56 | 29 s |
| LlamaParse Extract Cost-Effective | 49.6 | 83.5 | −33.9 | $8.18 | 29 s |
| Reducto Deep Extract | 49.4 | 85.6 | −36.2 | $40 | 65 s |
| Extend Max Context | 46.5 | 84.7 | −38.2 | $50 | 43 s |
| Reducto Extract | 39.4 | 81.7 | −42.3 | $20 | 16 s |
What to do about it
Every system struggles here. Use the best one below, or send these to a person.
Published examples
2 public resumes of this kind.