Gemma 4 26B A4B
Stored tensors can include buffers and repeated shared weights. A verified unique parameter count is unavailable. Checkpoint size is not the memory required to run the model.
GeekRadar / The comparison desk
COMPARE.
AI modelsGemma 4 26B A4B
AI model Released
- Context window
- 262,144 tokens
- Open weights
- Yes
Capability & preference
General capability · ECI
A cross-benchmark capability estimate. Index points, not a percentage; intervals show uncertainty.
Human preference · Arena
Overall text preferences from head-to-head votes. Not factual accuracy or a task-success rate.
Benchmark results
- AI2D_TEST · version unreported · reported score88.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- AIME26 · version unreported · reported score · Full AIME 2026 I and II88.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Task set
- Full AIME 2026 I and II*
- C-Eval · version unreported · reported score82.50*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- CC-OCR · version unreported · reported score74.50*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- CharXiv(RQ) · version unreported · reported score69*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Claw-Eval Avg · version unreported · reported score58.80*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Claw-Eval Pass^3 · version unreported · reported score28*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- DeepPlanning · version unreported · reported score16.20*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- GPQA · version unreported · reported score82.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HallusionBench · version unreported · reported score66.10*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HLE · version unreported · reported score8.70*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HMMT Feb 25 · version unreported · reported score91.70*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HMMT Feb 26 · version unreported · reported score79*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HMMT Nov 25 · version unreported · reported score87.50*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- IMOAnswerBench · version unreported · reported score74.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- LiveCodeBench v6 · 6 · reported score77.10*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Mathvista(mini) · version unreported · reported score79.40*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MCP-Atlas · version unreported · reported score · Public set50*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Judge
- gemini-2.5-pro*
- Task set
- Public set*
- MCPMark · version unreported · reported score · GitHub MCP v0.30.314.20*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- GitHub MCP v0.30.3*
- Truncation
- Playwright responses at 32K tokens*
- MMBench EN-DEV-v1.1 · 1.1 · reported score89*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MMLU-Pro · version unreported · reported score82.60*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MMLU-Redux · version unreported · reported score92.70*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MMMU · version unreported · reported score78.40*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MMMU-Pro · version unreported · reported score73.80*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Original footnote
- Asterisk in reported result; its qualification is not specified in the retained table*
- NL2Repo · version unreported · reported score · Claude Code11.60*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Claude Code*
- Variant
- Maximum 900 turns*
- Temperature
- 1.0*
- Top p
- 0.95*
- OmniDocBench1.5 · 1.5 · reported score74.40*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- QwenClawBench · version unreported · reported score · Internal real-user distribution38.70*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Task set
- Internal real-user distribution*
- Temperature
- 0.6*
- Context length
- 256K tokens*
- QwenWebBench · version unreported · reported score · Internal1,178* EloPlatform unreported Compare results
Test conditions
- Metric
- BT/Elo rating*
- Checkpoint version
- Not specified*
- Languages
- English and Chinese*
- Categories
- Seven frontend application categories*
- Judge
- Multimodal automatic judge*
- Task set
- Internal*
- RealWorldQA · version unreported · reported score72.20*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SimpleVQA · version unreported · reported score52.20*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SkillsBench Avg5 · version unreported · reported score · OpenCode · Self-contained subset, API-dependent tasks excluded · Mean of five runs12.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- OpenCode*
- Tasks
- 78*
- Runs
- 5*
- Statistic
- Mean of five runs*
- Task set
- Self-contained subset, API-dependent tasks excluded*
- SuperGPQA · version unreported · reported score61.40*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SWE-bench Multilingual · version unreported · reported score · Internal agent scaffold: bash and file editing · Public benchmark17.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Internal agent scaffold: bash and file editing*
- Temperature
- 1.0*
- Top p
- 0.95*
- Context length
- 200K tokens*
- Task set
- Public benchmark*
- SWE-bench Pro · version unreported · reported score · Internal agent scaffold: bash and file editing · Corrected public SWE-bench Pro task set13.80*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Internal agent scaffold: bash and file editing*
- Temperature
- 1.0*
- Top p
- 0.95*
- Context length
- 200K tokens*
- Task set
- Corrected public SWE-bench Pro task set*
- SWE-bench Verified · version unreported · reported score · Internal agent scaffold: bash and file editing · Public benchmark17.40*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Internal agent scaffold: bash and file editing*
- Temperature
- 1.0*
- Top p
- 0.95*
- Context length
- 200K tokens*
- Task set
- Public benchmark*
- TAU3-Bench · version unreported · reported score59*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- User model
- gpt-5.2, low reasoning effort*
- Retrieval
- Default BM25*
- Terminal-Bench 2.0 · 2.0 · reported score · Harbor/Terminus-2 · Mean of five runs34.20*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Harbor/Terminus-2*
- Timeout
- 3 hours*
- Hardware
- 32 CPU / 48 GB RAM*
- Runs
- 5*
- Statistic
- Mean of five runs*
- Temperature
- 1.0*
- Top p
- 0.95*
- Top k
- 20*
- Maximum output
- 80K tokens*
- Context length
- 256K tokens*
- Tool Decathlon · version unreported · reported score12*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- VideoMMMU · version unreported · reported score76*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- VITA-Bench · version unreported · reported score · Average subdomain score36.90*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Judge
- claude-4-sonnet*
- Statistic
- Average subdomain score*
- WideSearch · version unreported · reported score38.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- ZEROBench_sub · version unreported · reported score26.30*Platform unreported Compare results
Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
Recorded results
Only the same benchmark, version, workload, scenario and unit share a row. Missing results are not zero. System results include the full submitted hardware and software configuration.
Arena text preference / chinesehuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,495.50 rating
Confidence interval: 1,470.60 – 1,520.39
523 votes
- Category
- chinese
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / codinghuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,446.34 rating
Confidence interval: 1,431.67 – 1,461.01
1442 votes
- Category
- coding
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / creative_writinghuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,401.47 rating
Confidence interval: 1,383.28 – 1,419.66
1000 votes
- Category
- creative_writing
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / englishhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,443.51 rating
Confidence interval: 1,431.59 – 1,455.43
2247 votes
- Category
- english
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / exclude_tieshuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,428.84 rating
Confidence interval: 1,418.07 – 1,439.61
4166 votes
- Category
- exclude_ties
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / experthuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,447.33 rating
Confidence interval: 1,421.96 – 1,472.70
455 votes
- Category
- expert
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / frenchhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,461.66 rating
Confidence interval: 1,424.68 – 1,498.63
261 votes
- Category
- french
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / hard_promptshuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,438.51 rating
Confidence interval: 1,428.78 – 1,448.24
3470 votes
- Category
- hard_prompts
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / hard_prompts_englishhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,447.77 rating
Confidence interval: 1,432.00 – 1,463.53
1276 votes
- Category
- hard_prompts_english
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_business_and_management_and_financial_operationshuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,429.10 rating
Confidence interval: 1,412.20 – 1,446.01
1113 votes
- Category
- industry_business_and_management_and_financial_operations
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_entertainment_and_sports_and_mediahuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,403.54 rating
Confidence interval: 1,386.22 – 1,420.86
1111 votes
- Category
- industry_entertainment_and_sports_and_media
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_legal_and_governmenthuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,455.94 rating
Confidence interval: 1,428.71 – 1,483.17
415 votes
- Category
- industry_legal_and_government
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_life_and_physical_and_social_sciencehuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,446.43 rating
Confidence interval: 1,427.94 – 1,464.92
927 votes
- Category
- industry_life_and_physical_and_social_science
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_mathematicalhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,477.00 rating
Confidence interval: 1,445.74 – 1,508.26
309 votes
- Category
- industry_mathematical
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_medicine_and_healthcarehuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,427.58 rating
Confidence interval: 1,398.94 – 1,456.23
391 votes
- Category
- industry_medicine_and_healthcare
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_software_and_it_serviceshuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,453.64 rating
Confidence interval: 1,441.72 – 1,465.55
2211 votes
- Category
- industry_software_and_it_services
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / industry_writing_and_literature_and_languagehuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,405.87 rating
Confidence interval: 1,390.70 – 1,421.03
1395 votes
- Category
- industry_writing_and_literature_and_language
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / instruction_followinghuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,419.96 rating
Confidence interval: 1,406.51 – 1,433.41
1736 votes
- Category
- instruction_following
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / longer_queryhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,427.33 rating
Confidence interval: 1,413.62 – 1,441.05
1714 votes
- Category
- longer_query
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / mathhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,469.71 rating
Confidence interval: 1,442.13 – 1,497.30
389 votes
- Category
- math
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / multi_turnhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,441.59 rating
Confidence interval: 1,424.65 – 1,458.52
1143 votes
- Category
- multi_turn
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / non_englishhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,421.72 rating
Confidence interval: 1,412.46 – 1,430.97
3843 votes
- Category
- non_english
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / overallhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,434.17 rating
Confidence interval: 1,426.75 – 1,441.58
6090 votes
- Category
- overall
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / russianhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,434.16 rating
Confidence interval: 1,416.14 – 1,452.18
900 votes
- Category
- russian
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / spanishhuman preference · 2026-10-28 · rating
Gemma 4 26B A4B
1,417.11 rating
Confidence interval: 1,375.46 – 1,458.77
202 votes
- Category
- spanish
- Evaluated configuration
- Gemma 4 26B A4B
- Evaluation type
- human preference
Arena text preference / chinesehuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,495.41 rating
Confidence interval: 1,470.56 – 1,520.27
524 votes
- Category
- chinese
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / codinghuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,446.43 rating
Confidence interval: 1,431.76 – 1,461.09
1445 votes
- Category
- coding
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / creative_writinghuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,401.45 rating
Confidence interval: 1,383.28 – 1,419.62
1001 votes
- Category
- creative_writing
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / englishhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,443.91 rating
Confidence interval: 1,431.99 – 1,455.83
2249 votes
- Category
- english
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / exclude_tieshuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,428.76 rating
Confidence interval: 1,417.99 – 1,439.53
4167 votes
- Category
- exclude_ties
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / experthuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,447.27 rating
Confidence interval: 1,421.89 – 1,472.65
455 votes
- Category
- expert
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / frenchhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,459.68 rating
Confidence interval: 1,422.73 – 1,496.64
262 votes
- Category
- french
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / hard_promptshuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,438.48 rating
Confidence interval: 1,428.76 – 1,448.21
3473 votes
- Category
- hard_prompts
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / hard_prompts_englishhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,448.14 rating
Confidence interval: 1,432.37 – 1,463.91
1277 votes
- Category
- hard_prompts_english
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_business_and_management_and_financial_operationshuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,429.22 rating
Confidence interval: 1,412.34 – 1,446.10
1115 votes
- Category
- industry_business_and_management_and_financial_operations
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_entertainment_and_sports_and_mediahuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,403.60 rating
Confidence interval: 1,386.28 – 1,420.93
1111 votes
- Category
- industry_entertainment_and_sports_and_media
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_legal_and_governmenthuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,455.95 rating
Confidence interval: 1,428.71 – 1,483.18
415 votes
- Category
- industry_legal_and_government
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_life_and_physical_and_social_sciencehuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,446.46 rating
Confidence interval: 1,427.99 – 1,464.93
928 votes
- Category
- industry_life_and_physical_and_social_science
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_mathematicalhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,475.38 rating
Confidence interval: 1,444.23 – 1,506.53
311 votes
- Category
- industry_mathematical
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_medicine_and_healthcarehuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,427.64 rating
Confidence interval: 1,398.99 – 1,456.29
391 votes
- Category
- industry_medicine_and_healthcare
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_software_and_it_serviceshuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,453.61 rating
Confidence interval: 1,441.70 – 1,465.52
2214 votes
- Category
- industry_software_and_it_services
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / industry_writing_and_literature_and_languagehuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,405.60 rating
Confidence interval: 1,390.44 – 1,420.77
1395 votes
- Category
- industry_writing_and_literature_and_language
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / instruction_followinghuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,420.14 rating
Confidence interval: 1,406.70 – 1,433.59
1738 votes
- Category
- instruction_following
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / longer_queryhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,427.27 rating
Confidence interval: 1,413.57 – 1,440.98
1715 votes
- Category
- longer_query
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / mathhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,469.69 rating
Confidence interval: 1,442.17 – 1,497.21
390 votes
- Category
- math
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / multi_turnhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,441.33 rating
Confidence interval: 1,424.40 – 1,458.26
1143 votes
- Category
- multi_turn
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / non_englishhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,421.43 rating
Confidence interval: 1,412.19 – 1,430.68
3845 votes
- Category
- non_english
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / overallhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,434.12 rating
Confidence interval: 1,426.70 – 1,441.53
6094 votes
- Category
- overall
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / russianhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,433.52 rating
Confidence interval: 1,415.49 – 1,451.56
898 votes
- Category
- russian
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Arena text preference / spanishhuman preference · 2026-10-15 · rating
Gemma 4 26B A4B
1,417.07 rating
Confidence interval: 1,375.60 – 1,458.55
203 votes
- Category
- spanish
- Evaluated configuration
- gemma-4-26b-a4b
- Evaluation type
- human preference
Run it locally
Capabilities
- Context window
- 262,144 tokens
- Open weights
- Yes
- Checkpoint tensor payload
- 48.07 GiB
Input modalities
- text
- image
Output modalities
- text
Processing & platform
- Architecture
- gemma4
System configuration
Detailed specifications
Specifications · 42 fields
- NL2Repo
- suite
- NL2Repo
- workload
- reported score · Claude Code
- score
- 11.6*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Claude Code*
- Variant
- Maximum 900 turns*
- Temperature
- 1.0*
- Top p
- 0.95*
- HallusionBench
- suite
- HallusionBench
- workload
- reported score
- score
- 66.1*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- QwenWebBench
- suite
- QwenWebBench
- workload
- reported score · Internal
- score
- 1,178.0*
- unit
- Elo
Scope and conditions
- Test conditions
- Metric
- BT/Elo rating*
- Checkpoint version
- Not specified*
- Languages
- English and Chinese*
- Categories
- Seven frontend application categories*
- Judge
- Multimodal automatic judge*
- Task set
- Internal*
- OmniDocBench1.5
- suite
- OmniDocBench1.5
- version
- 1.5
- workload
- reported score
- score
- 74.4*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- TAU3-Bench
- suite
- TAU3-Bench
- workload
- reported score
- score
- 59.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- User model
- gpt-5.2, low reasoning effort*
- Retrieval
- Default BM25*
- CharXiv(RQ)
- suite
- CharXiv(RQ)
- workload
- reported score
- score
- 69.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- VITA-Bench
- suite
- VITA-Bench
- workload
- reported score · Average subdomain score
- score
- 36.9*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Judge
- claude-4-sonnet*
- Statistic
- Average subdomain score*
- CC-OCR
- suite
- CC-OCR
- workload
- reported score
- score
- 74.5*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- DeepPlanning
- suite
- DeepPlanning
- workload
- reported score
- score
- 16.2*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- AI2D TEST
- suite
- AI2D_TEST
- workload
- reported score
- score
- 88.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Tool Decathlon
- suite
- Tool Decathlon
- workload
- reported score
- score
- 12.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MCPMark
- suite
- MCPMark
- workload
- reported score · GitHub MCP v0.30.3
- score
- 14.2*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- GitHub MCP v0.30.3*
- Truncation
- Playwright responses at 32K tokens*
- MCP-Atlas
- suite
- MCP-Atlas
- workload
- reported score · Public set
- score
- 50.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Judge
- gemini-2.5-pro*
- Task set
- Public set*
- WideSearch
- suite
- WideSearch
- workload
- reported score
- score
- 38.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SWE-bench Verified
- suite
- SWE-bench Verified
- workload
- reported score · Internal agent scaffold: bash and file editing · Public benchmark
- score
- 17.4*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Internal agent scaffold: bash and file editing*
- Temperature
- 1.0*
- Top p
- 0.95*
- Context length
- 200K tokens*
- Task set
- Public benchmark*
- MMMU
- suite
- MMMU
- workload
- reported score
- score
- 78.4*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MMLU-Pro
- suite
- MMLU-Pro
- workload
- reported score
- score
- 82.6*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- MMLU-Redux
- suite
- MMLU-Redux
- workload
- reported score
- score
- 92.7*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SuperGPQA
- suite
- SuperGPQA
- workload
- reported score
- score
- 61.4*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- VideoMMMU
- suite
- VideoMMMU
- workload
- reported score
- score
- 76.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- C-Eval
- suite
- C-Eval
- workload
- reported score
- score
- 82.5*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- GPQA
- suite
- GPQA
- workload
- reported score
- score
- 82.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HLE
- suite
- HLE
- workload
- reported score
- score
- 8.7*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- LiveCodeBench v6
- suite
- LiveCodeBench v6
- version
- 6
- workload
- reported score
- score
- 77.1*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HMMT Feb 25
- suite
- HMMT Feb 25
- workload
- reported score
- score
- 91.7*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SWE-bench Multilingual
- suite
- SWE-bench Multilingual
- workload
- reported score · Internal agent scaffold: bash and file editing · Public benchmark
- score
- 17.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Internal agent scaffold: bash and file editing*
- Temperature
- 1.0*
- Top p
- 0.95*
- Context length
- 200K tokens*
- Task set
- Public benchmark*
- MMMU-Pro
- suite
- MMMU-Pro
- workload
- reported score
- score
- 73.8*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Original footnote
- Asterisk in reported result; its qualification is not specified in the retained table*
- HMMT Nov 25
- suite
- HMMT Nov 25
- workload
- reported score
- score
- 87.5*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- HMMT Feb 26
- suite
- HMMT Feb 26
- workload
- reported score
- score
- 79.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- IMOAnswerBench
- suite
- IMOAnswerBench
- workload
- reported score
- score
- 74.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- AIME26
- suite
- AIME26
- workload
- reported score · Full AIME 2026 I and II
- score
- 88.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Task set
- Full AIME 2026 I and II*
- Mathvista(mini)
- suite
- Mathvista(mini)
- workload
- reported score
- score
- 79.4*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SWE-bench Pro
- suite
- SWE-bench Pro
- workload
- reported score · Internal agent scaffold: bash and file editing · Corrected public SWE-bench Pro task set
- score
- 13.8*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Internal agent scaffold: bash and file editing*
- Temperature
- 1.0*
- Top p
- 0.95*
- Context length
- 200K tokens*
- Task set
- Corrected public SWE-bench Pro task set*
- Terminal-Bench 2.0
- suite
- Terminal-Bench 2.0
- version
- 2.0
- workload
- reported score · Harbor/Terminus-2 · Mean of five runs
- score
- 34.2*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- Harbor/Terminus-2*
- Timeout
- 3 hours*
- Hardware
- 32 CPU / 48 GB RAM*
- Runs
- 5*
- Statistic
- Mean of five runs*
- Temperature
- 1.0*
- Top p
- 0.95*
- Top k
- 20*
- Maximum output
- 80K tokens*
- Context length
- 256K tokens*
- ZEROBench sub
- suite
- ZEROBench_sub
- workload
- reported score
- score
- 26.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Claw-Eval Avg
- suite
- Claw-Eval Avg
- workload
- reported score
- score
- 58.8*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- RealWorldQA
- suite
- RealWorldQA
- workload
- reported score
- score
- 72.2*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Claw-Eval Pass^3
- suite
- Claw-Eval Pass^3
- workload
- reported score
- score
- 28.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SkillsBench Avg5
- suite
- SkillsBench Avg5
- workload
- reported score · OpenCode · Self-contained subset, API-dependent tasks excluded · Mean of five runs
- score
- 12.3*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Harness
- OpenCode*
- Tasks
- 78*
- Runs
- 5*
- Statistic
- Mean of five runs*
- Task set
- Self-contained subset, API-dependent tasks excluded*
- MMBench EN-DEV-v1.1
- suite
- MMBench EN-DEV-v1.1
- version
- 1.1
- workload
- reported score
- score
- 89.0*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- SimpleVQA
- suite
- SimpleVQA
- workload
- reported score
- score
- 52.2*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- QwenClawBench
- suite
- QwenClawBench
- workload
- reported score · Internal real-user distribution
- score
- 38.7*
Scope and conditions
- Test conditions
- Metric
- reported score*
- Checkpoint version
- Not specified*
- Task set
- Internal real-user distribution*
- Temperature
- 0.6*
- Context length
- 256K tokens*
0 selected Compare 0
* Provisional data