Alibaba · Vision LLM
Qwen 3.5 35B on ResumeBench: 77.9 F1, rank 14 of 18
Qwen 3.5 35B ranks 14th of 18 on ResumeBench: 77.9 F1, $3.91 per 1,000 pages, 78 seconds per resume. It is 9.2 points behind LlamaParse Extract Agentic Plus. Qwen 3.8 Flash scores 84.1 for $2.98.
7th of 11 vision LLMs. Strongest on content as images (+2.1 points), handwritten (+1.3 points) and table grid (+0.4 points). Weakest on two-column (−9.4 points), three-column (−9.2 points) and two-column sidebar (−8.5 points).
- Rank
- 14/18
- F1
- 77.9
- Precision
- 78.8
- Recall
- 83.7
- Perfect resumes
- 4/300
- Per 1,000 pages
- $3.91
- Median per resume
- 78 s
Averaged over 300 resumes. Dates (scored separately): 88.5. 90% finish within 221 s.
Where Qwen 3.5 35B wins and loses
F1 by kind of resume, against the median system.
Reading the page
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Handwritten | 9 | 52.7 | 51.5 | +1.3 | 7th |
| Image only | 28 | 69.7 | 71.4 | −1.7 | 12th |
| Sidebar or grid | 82 | 74.8 | 81.9 | −7.2 | 14th |
| Photo | 45 | 80.0 | 84.0 | −4.1 | 13th |
Distractors
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Rotated & warped pages | 21 | 76.5 | 82.9 | −6.4 | 12th |
| Hidden text & prompt injections | 14 | 79.2 | 84.5 | −5.3 | 12th |
| Content as images | 8 | 90.3 | 88.2 | +2.1 | 5th |
| Wording traps | 16 | 84.8 | 84.5 | +0.3 | 9th |
| Typos kept on purpose | 37 | 80.4 | 83.7 | −3.3 | 13th |
Extraction challenges
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Many entries | 83 | 78.9 | 81.9 | −2.9 | 14th |
| Long bullet lists | 101 | 76.3 | 81.4 | −5.1 | 14th |
| Split across pages | 85 | 76.5 | 82.2 | −5.7 | 14th |
| Big skills list | 60 | 76.6 | 83.6 | −7.0 | 14th |
| Custom sections | 105 | 77.0 | 80.8 | −3.8 | 14th |
| Combined headings | 37 | 74.6 | 82.7 | −8.1 | 14th |
| Promotions | 88 | 76.7 | 83.5 | −6.8 | 14th |
| Partial dates | 138 | 78.9 | 83.4 | −4.6 | 14th |
| Overlapping jobs | 69 | 76.2 | 81.7 | −5.5 | 14th |
| Accented text | 13 | 80.4 | 85.5 | −5.1 | 15th |
Noise
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Repeated header | 67 | 78.6 | 85.2 | −6.6 | 14th |
| References | 55 | 81.8 | 85.2 | −3.4 | 14th |
| Busy page | 18 | 71.8 | 80.1 | −8.3 | 14th |
Source
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Everyday resumes | 212 | 77.7 | 85.1 | −7.5 | 14th |
| Expert-built (Canva / LaTeX) | 32 | 83.7 | 83.6 | +0.0 | 9th |
| Resume-builder templates | 10 | 79.1 | 81.2 | −2.1 | 13th |
| Scanned & rotated | 19 | 75.3 | 82.9 | −7.6 | 12th |
| Distractor layouts | 18 | 85.7 | 85.3 | +0.4 | 7th |
Layout
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Single column | 195 | 79.4 | 84.0 | −4.6 | 14th |
| Two-column | 23 | 77.4 | 86.8 | −9.4 | 15th |
| Two-column sidebar | 54 | 71.5 | 80.0 | −8.5 | 14th |
| Three-column | 20 | 75.6 | 84.8 | −9.2 | 15th |
| Table grid | 8 | 94.5 | 94.1 | +0.4 | 9th |
Length
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Short | 148 | 79.8 | 87.1 | −7.3 | 15th |
| Medium | 90 | 79.7 | 84.7 | −4.9 | 14th |
| Long | 62 | 71.0 | 74.6 | −3.7 | 14th |
By section of the resume
F1 by resume section, against the median system.
| Section | Resumes | F1 | Median | Gap |
|---|---|---|---|---|
| Contact details | 300 | 92.1 | 93.5 | −1.4 |
| Experience | 300 | 79.3 | 86.3 | −7.1 |
| Education | 297 | 81.8 | 89.4 | −7.6 |
| Summary | 291 | 80.4 | 81.7 | −1.3 |
| Certifications | 251 | 79.1 | 81.1 | −2.0 |
| Awards | 116 | 61.9 | 75.4 | −13.5 |
| Skills | 294 | 87.8 | 92.1 | −4.3 |
| Languages | 175 | 79.4 | 84.3 | −4.9 |
| Custom sections | 236 | 40.2 | 49.9 | −9.7 |
| Interests | 66 | 89.7 | 91.7 | −2.0 |
| Volunteering | 80 | 83.1 | 84.5 | −1.4 |
| Profiles | 39 | 65.8 | 66.7 | −0.9 |
| Projects | 81 | 76.7 | 83.2 | −6.4 |
| Publications | 53 | 75.0 | 78.6 | −3.6 |
On the published resumes
Best and worst of the 20 public resumes.
| Resume | What makes it hard | F1 |
|---|---|---|
| Head of Developer Experience (Internal title: Jedi Master of Analytics) | Unusual job titles | 90.6 |
| Quality Systems Lead | Photo of a printout | 88.9 |
| Head of Events & Public Space Operations | Table grid with emoji bullets | 88.2 |
| Senior Game Designer | Game designer's Canva resume | 66.8 |
| senior Consultant - package Implementation | Handwritten resume | 59.2 |
| Senior ocularist and practice owner | Handwritten resume, six pages | 38.5 |
How it was run
Qwen3.5-35B-A3B (open weights), page images in, JSON schema out. Same PDF, schema and instructions as every other system, one run per resume (4 November 2026).
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 testClosest alternatives
Nearest in F1.
| Model | F1 | Per 1,000 pages | Median time | Head to head |
|---|---|---|---|---|
| GPT-5.5 | 83.5 | $87 | 48 s | Qwen 3.5 35B vs GPT-5.5 |
| Reducto Extract | 81.7 | $20 | 16 s | Qwen 3.5 35B vs Reducto Extract |
| Extend Extract | 81.0 | $12 | 31 s | Qwen 3.5 35B vs Extend Extract |
| Qwen 3.5 9B | 76.0 | $1.31 | 138 s | Qwen 3.5 35B vs Qwen 3.5 9B |
Qwen 3.5 35B against every other model
- Qwen 3.5 35B vs LlamaParse Extract Agentic Plus
- Qwen 3.5 35B vs LlamaParse Extract Agentic
- Qwen 3.5 35B vs LlamaParse Extract Cost-Effective
- Qwen 3.5 35B vs Reducto Deep Extract
- Qwen 3.5 35B vs Extend Max Context
- Qwen 3.5 35B vs GPT-5.6
- Qwen 3.5 35B vs GPT-5.4 Nano
- Qwen 3.5 35B vs Claude Opus 4.8
- Qwen 3.5 35B vs Gemini 3.5 Flash
- Qwen 3.5 35B vs Gemma 4 26B
- Qwen 3.5 35B vs Qwen 3.8 Flash
- Qwen 3.5 35B vs Qwen 3.8 27B
- Qwen 3.5 35B vs GLM 5.3 Flash
Embed this result
Shows its current rank. Updates with each run.