Skip to content

Which resume extraction model handles scanned and image-only resumes best?

28 of the 300 resumes are scanned and image-only resumes. The typical system scores 10.5 points lower on them. Best: LlamaParse Extract Agentic Plus, 84.2 F1.

PDFs with no text layer: scans, photos of printouts and exported images, where the words have to be read from pixels. Best document API: LlamaParse Extract Agentic Plus (84.2). Best vision LLM: Claude Opus 4.8 (77.0). LlamaParse Extract Cost-Effective ranks 4 places higher here than overall.

Every system on scanned and image-only resumes

F1 on the 28 scanned and image-only resumes, against each system’s F1 on all 300 resumes. Cost and time are for all resumes.

Ranking on scanned and image-only resumes
ModelF1 hereF1 overallGapPer 1,000 pagesMedian time
LlamaParse Extract Agentic Plus84.287.2−3.0$5470 s
LlamaParse Extract Agentic81.585.5−4.1$2355 s
Claude Opus 4.877.085.5−8.5$9031 s
Extend Max Context76.584.7−8.3$5043 s
Reducto Deep Extract76.485.6−9.1$4065 s
LlamaParse Extract Cost-Effective76.383.5−7.2$8.1829 s
Gemini 3.5 Flash75.185.5−10.3$3543 s
Qwen 3.8 Flash74.784.1−9.4$2.98122 s
Reducto Extract72.081.7−9.7$2016 s
Extend Extract70.781.0−10.2$1231 s
Qwen 3.8 27B70.583.8−13.2$9.5936 s
Qwen 3.5 35B69.777.9−8.2$3.9178 s
GPT-5.665.484.7−19.3$7438 s
Qwen 3.5 9B61.276.0−14.8$1.31138 s
GPT-5.557.083.5−26.5$8748 s
GLM 5.3 Flash53.359.2−6.0$0.8046 s
Gemma 4 26B45.162.8−17.7$0.5629 s
GPT-5.4 Nano21.061.9−40.9$2.7620 s

What to do about it

If you get a lot of scans, rank parsers on this table, not the overall one.

Published examples

4 public resumes of this kind.

Other hard resumes

Frequently asked questions (FAQs)

Which resume extraction model is best for scanned and image-only resumes?

LlamaParse Extract Agentic Plus, at 84.2 F1.

How hard are scanned and image-only resumes for parsers?

The typical model scores 10.5 points lower on them than on other resumes.

How many scanned and image-only resumes are in ResumeBench?

28 of the 300.