BlogLLM Benchmark Scores as Published, Side by Side
Gemma 4 12B: Benchmark Scores as Published
Every result the catalog holds for Gemma 4 12B: 16 tests in 8 areas, from 1 page. All were printed by Google itself. Each number is as printed; nothing is ranked or rescored.
Developer GoogleReleased 2026-04-02Context window 256KInput Text, image, audioOutput TextWeights Open (Apache 2.0)Data version 2026103001, as of 2026-10-30
Science and Math
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| GPQA Diamond | 78.8% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
| AIME 2026 | 77.5% | Google, Gemma 4 model cardchecked 2026-10-15 | Tools: none |
Long Context
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| MRCR v2 8-needle 128K | 43.4% | Google, Gemma 4 model cardchecked 2026-10-15 | Context: 128k (average) |
Multimodal Understanding
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| MMMU-Pro | 69.1% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
| MathVision | 79.7% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
| OmniDocBench 1.5 (overall edit distance)lower is better | 0.164 | Google, Gemma 4 model cardchecked 2026-10-15 | - |
| CoVoST (metric not stated) | 38.5 | Google, Gemma 4 model cardchecked 2026-10-15 | - |
| FLEURS (metric not stated)lower is better | 0.069 | Google, Gemma 4 model cardchecked 2026-10-15 | - |
General Knowledge and Reasoning
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| Humanity's Last Exam (no tools) | 5.2% | Google, Gemma 4 model cardchecked 2026-10-15 | Tools: none |
| MMLU-Pro | 77.2% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
| BIG-Bench Extra Hard | 53% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
Code Generation
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| LiveCodeBench v6 | 72% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
| Codeforces rating | 1659 | Google, Gemma 4 model cardchecked 2026-10-15 | - |
Tool Use and Workflows
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| τ²-bench (average of 3 domains) | 69% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
Multilingual
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| MMMLU | 83.4% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
Healthcare
| Test | Value | Printed by | Setup as printed |
|---|---|---|---|
| MedXpertQA MM | 48.7% | Google, Gemma 4 model cardchecked 2026-10-15 | - |
More from Google
Every number here is read from the same catalog file as the benchmark table, data version 2026103001. Where one page printed the model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.