Skip to content

Alibaba · Vision LLM

Qwen 3.8 27B on ResumeBench: 83.8 F1, rank 9 of 18

Qwen 3.8 27B ranks 9th of 18 on ResumeBench: 83.8 F1, $9.59 per 1,000 pages, 36 seconds per resume. It is 3.4 points behind LlamaParse Extract Agentic Plus. Qwen 3.8 Flash scores 84.1 for $2.98.

5th of 11 vision LLMs. Strongest on resume-builder templates (+5.5 points), wording traps (+3.1 points) and distractor layouts (+2.3 points). Weakest on scanned & rotated (−4.5 points), rotated & warped pages (−4.2 points) and hidden text & prompt injections (−2.2 points).

Rank
9/18
F1
83.8
Precision
85.4
Recall
85.5
Perfect resumes
31/300
Per 1,000 pages
$9.59
Median per resume
36 s

Averaged over 300 resumes. Dates (scored separately): 91.6. 90% finish within 167 s.

Where Qwen 3.8 27B wins and loses

F1 by kind of resume, against the median system.

Reading the page

Qwen 3.8 27B by reading the page
Kind of resume# of resumesF1MedianGapRank
Handwritten951.651.5+0.19th
Image only2870.571.4−0.811th
Sidebar or grid8281.881.9−0.210th
Photo4584.184.0+0.09th

Distractors

Qwen 3.8 27B by distractors
Kind of resume# of resumesF1MedianGapRank
Rotated & warped pages2178.782.9−4.211th
Hidden text & prompt injections1482.484.5−2.211th
Content as images889.588.2+1.36th
Wording traps1687.684.5+3.14th
Typos kept on purpose3782.983.7−0.710th

Extraction challenges

Qwen 3.8 27B by extraction challenges
Kind of resume# of resumesF1MedianGapRank
Many entries8381.681.9−0.210th
Long bullet lists10180.881.4−0.610th
Split across pages8581.882.2−0.410th
Big skills list6082.983.6−0.812th
Custom sections10581.080.8+0.28th
Combined headings3783.782.7+1.05th
Promotions8883.283.5−0.310th
Partial dates13883.683.4+0.19th
Overlapping jobs6982.481.7+0.68th
Accented text1384.985.5−0.611th

Noise

Qwen 3.8 27B by noise
Kind of resume# of resumesF1MedianGapRank
Repeated header6785.285.2+0.19th
References5584.985.2−0.210th
Busy page1880.480.1+0.39th

Source

Qwen 3.8 27B by source
Kind of resume# of resumesF1MedianGapRank
Everyday resumes21285.285.1+0.19th
Expert-built (Canva / LaTeX)3283.483.6−0.211th
Resume-builder templates1086.781.2+5.51st
Scanned & rotated1978.482.9−4.511th
Distractor layouts1887.685.3+2.34th

Layout

Qwen 3.8 27B by layout
Kind of resume# of resumesF1MedianGapRank
Single column19584.184.0+0.19th
Two-column2387.986.8+1.17th
Two-column sidebar5478.480.0−1.610th
Three-column2085.584.8+0.77th
Table grid895.794.1+1.63rd

Length

Qwen 3.8 27B by length
Kind of resume# of resumesF1MedianGapRank
Short14887.187.1−0.010th
Medium9085.584.7+0.85th
Long6273.474.6−1.212th

By section of the resume

F1 by resume section, against the median system.

Qwen 3.8 27B by resume section
SectionResumesF1MedianGap
Contact details30093.293.5−0.4
Experience30086.286.3−0.2
Education29787.289.4−2.3
Summary29081.781.7+0.0
Certifications24482.781.1+1.6
Awards9778.475.4+3.0
Skills28492.792.1+0.6
Languages16885.484.3+1.0
Custom sections21651.249.9+1.3
Interests6394.691.7+2.9
Volunteering7984.784.5+0.1
Profiles3472.566.7+5.8
Projects7885.583.2+2.3
Publications5177.278.6−1.3

On the published resumes

Best and worst of the 20 public resumes.

Qwen 3.8 27B on published resumes
ResumeWhat makes it hardF1
Head of Developer Experience (Internal title: Jedi Master of Analytics) Unusual job titles96.3
Supervisory Human Resources Specialist, GS-0201-15Resume-builder template88.5
Head of Events & Public Space OperationsTable grid with emoji bullets88.1
Deputy Head of MissionRotated scan with hidden identity text65.6
senior Consultant - package ImplementationHandwritten resume46.5
Senior ocularist and practice ownerHandwritten resume, six pages41.2

How it was run

Qwen3.8-27B (open weights), page images in, JSON schema out. Same PDF, schema and instructions as every other system, one run per resume (4 November 2026).

Qwen 3.8 27B on Alibaba’s site

Run the open-source harness: list its pipelines, or grade your own predictions (GitHub):

$ pip install "resume-bench[runners] @ git+https://github.com/careerflow/ResumeBench"
$ hf auth login   # the dataset is gated: accept its terms on Hugging Face first
$ resume-bench download
$ resume-bench providers
$ resume-bench grade-file my_predictions.jsonl --split test

Closest alternatives

Nearest in F1.

Embed this result

Shows its current rank. Updates with each run.

Qwen 3.8 27B: #9 on ResumeBench resume extraction benchmark
<a href="https://www.resumebench.ai/systems/qwen-3-8-27b"><img src="https://www.resumebench.ai/badge/qwen-3-8-27b.svg" alt="Qwen 3.8 27B: #9 on ResumeBench resume extraction benchmark" height="28"></a>

Choosing a resume extraction model

All 18 models · Leaderboard

Frequently asked questions (FAQs)

How accurate is Qwen 3.8 27B at resume extraction?

83.8 F1 on ResumeBench's 300 resumes, 9th of 18 models.

How much does Qwen 3.8 27B cost to parse resumes?

$9.59 per 1,000 pages, about $0.027 per resume.

How fast is Qwen 3.8 27B?

36 seconds per resume (median).

Which resumes is Qwen 3.8 27B best at?

Resume-builder templates, wording traps and distractor layouts. Weakest on scanned & rotated, rotated & warped pages and hidden text & prompt injections.

What is a cheaper alternative to Qwen 3.8 27B?

Qwen 3.8 Flash: 84.1 F1 for $2.98 per 1,000 pages.