Google · Vision LLM
Gemma 4 26B on ResumeBench: 62.8 F1, rank 16 of 18
Gemma 4 26B ranks 16th of 18 on ResumeBench: 62.8 F1, $0.56 per 1,000 pages, 29 seconds per resume. It is 24.3 points behind LlamaParse Extract Agentic Plus.
9th of 11 vision LLMs. Strongest on handwritten (−1.2 points), expert-built (canva / latex) (−13.4 points) and table grid (−14.2 points). Weakest on scanned & rotated (−42.2 points), rotated & warped pages (−39.5 points) and hidden text & prompt injections (−36.0 points).
- Rank
- 16/18
- F1
- 62.8
- Precision
- 69.1
- Recall
- 66.2
- Perfect resumes
- 4/300
- Per 1,000 pages
- $0.56
- Median per resume
- 29 s
Averaged over 300 resumes. Dates (scored separately): 82.5. 90% finish within 76 s.
Where Gemma 4 26B 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 | 50.3 | 51.5 | −1.2 | 14th |
| Image only | 28 | 45.1 | 71.4 | −26.2 | 17th |
| Sidebar or grid | 82 | 60.4 | 81.9 | −21.5 | 17th |
| Photo | 45 | 66.9 | 84.0 | −17.2 | 16th |
Distractors
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Rotated & warped pages | 21 | 43.4 | 82.9 | −39.5 | 17th |
| Hidden text & prompt injections | 14 | 48.6 | 84.5 | −36.0 | 16th |
| Content as images | 8 | 65.6 | 88.2 | −22.6 | 16th |
| Wording traps | 16 | 60.6 | 84.5 | −23.9 | 17th |
| Typos kept on purpose | 37 | 51.3 | 83.7 | −32.4 | 16th |
Extraction challenges
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Many entries | 83 | 58.0 | 81.9 | −23.8 | 17th |
| Long bullet lists | 101 | 57.4 | 81.4 | −24.0 | 17th |
| Split across pages | 85 | 60.0 | 82.2 | −22.2 | 17th |
| Big skills list | 60 | 61.9 | 83.6 | −21.7 | 18th |
| Custom sections | 105 | 58.9 | 80.8 | −21.9 | 17th |
| Combined headings | 37 | 60.9 | 82.7 | −21.8 | 18th |
| Promotions | 88 | 59.5 | 83.5 | −24.0 | 16th |
| Partial dates | 138 | 62.9 | 83.4 | −20.5 | 16th |
| Overlapping jobs | 69 | 57.5 | 81.7 | −24.3 | 17th |
| Accented text | 13 | 59.3 | 85.5 | −26.2 | 16th |
Noise
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Repeated header | 67 | 64.0 | 85.2 | −21.2 | 18th |
| References | 55 | 62.5 | 85.2 | −22.6 | 17th |
| Busy page | 18 | 61.5 | 80.1 | −18.6 | 17th |
Source
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Everyday resumes | 212 | 64.5 | 85.1 | −20.7 | 18th |
| Expert-built (Canva / LaTeX) | 32 | 70.2 | 83.6 | −13.4 | 16th |
| Resume-builder templates | 10 | 58.6 | 81.2 | −22.6 | 17th |
| Scanned & rotated | 19 | 40.7 | 82.9 | −42.2 | 17th |
| Distractor layouts | 18 | 62.4 | 85.3 | −23.0 | 16th |
Layout
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Single column | 195 | 63.1 | 84.0 | −20.9 | 17th |
| Two-column | 23 | 69.3 | 86.8 | −17.5 | 17th |
| Two-column sidebar | 54 | 55.1 | 80.0 | −24.9 | 17th |
| Three-column | 20 | 67.0 | 84.8 | −17.8 | 17th |
| Table grid | 8 | 79.8 | 94.1 | −14.2 | 17th |
Length
| Kind of resume | # of resumes | F1 | Median | Gap | Rank |
|---|---|---|---|---|---|
| Short | 148 | 65.8 | 87.1 | −21.3 | 17th |
| Medium | 90 | 63.2 | 84.7 | −21.5 | 16th |
| Long | 62 | 55.1 | 74.6 | −19.6 | 18th |
By section of the resume
F1 by resume section, against the median system.
| Section | Resumes | F1 | Median | Gap |
|---|---|---|---|---|
| Contact details | 300 | 84.6 | 93.5 | −9.0 |
| Experience | 300 | 56.0 | 86.3 | −30.4 |
| Education | 297 | 68.0 | 89.4 | −21.4 |
| Summary | 290 | 35.5 | 81.7 | −46.2 |
| Certifications | 244 | 63.0 | 81.1 | −18.1 |
| Awards | 110 | 55.7 | 75.4 | −19.7 |
| Skills | 290 | 81.9 | 92.1 | −10.2 |
| Languages | 171 | 80.9 | 84.3 | −3.4 |
| Custom sections | 223 | 32.3 | 49.9 | −17.6 |
| Interests | 67 | 71.8 | 91.7 | −19.9 |
| Volunteering | 78 | 59.4 | 84.5 | −25.1 |
| Profiles | 37 | 57.2 | 66.7 | −9.5 |
| Projects | 83 | 55.0 | 83.2 | −28.2 |
| Publications | 53 | 54.5 | 78.6 | −24.1 |
On the published resumes
Best and worst of the 20 public resumes.
| Resume | What makes it hard | F1 |
|---|---|---|
| Senior Customer Service Representative | Busy Canva page | 80.2 |
| Head of Events & Public Space Operations | Table grid with emoji bullets | 78.0 |
| Senior Data Scientist (Lead) | Fifty skills, seven pages | 70.7 |
| Senior ocularist and practice owner | Handwritten resume, six pages | 37.0 |
| Deputy Head of Mission | Rotated scan with hidden identity text | 34.1 |
| Supervisory Human Resources Specialist, GS-0201-15 | Resume-builder template | 29.4 |
How it was run
Gemma 4 26B A4B (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 |
|---|---|---|---|---|
| Qwen 3.5 35B | 77.9 | $3.91 | 78 s | Gemma 4 26B vs Qwen 3.5 35B |
| Qwen 3.5 9B | 76.0 | $1.31 | 138 s | Gemma 4 26B vs Qwen 3.5 9B |
| GPT-5.4 Nano | 61.9 | $2.76 | 20 s | Gemma 4 26B vs GPT-5.4 Nano |
| GLM 5.3 Flash | 59.2 | $0.80 | 46 s | Gemma 4 26B vs GLM 5.3 Flash |
Gemma 4 26B against every other model
- Gemma 4 26B vs LlamaParse Extract Agentic Plus
- Gemma 4 26B vs LlamaParse Extract Agentic
- Gemma 4 26B vs LlamaParse Extract Cost-Effective
- Gemma 4 26B vs Reducto Deep Extract
- Gemma 4 26B vs Reducto Extract
- Gemma 4 26B vs Extend Max Context
- Gemma 4 26B vs Extend Extract
- Gemma 4 26B vs GPT-5.6
- Gemma 4 26B vs GPT-5.5
- Gemma 4 26B vs Claude Opus 4.8
- Gemma 4 26B vs Gemini 3.5 Flash
- Gemma 4 26B vs Qwen 3.8 Flash
- Gemma 4 26B vs Qwen 3.8 27B
Embed this result
Shows its current rank. Updates with each run.