← Overview
gemma-4-12b-it
- Quantization
- Q8
- Reasoning
- off
- Profile
- quick
- Suite
- v2b
- Date
- 2026-09-18 – 2026-09-19
- Hardware
- NVIDIA DGX Spark, 128 GB RAM
- Inference software
- vLLM 0.26.0
- Measured
- directly at the inference server
- Overall
- 65.8
Categories
| Category | Score |
|---|
| Knowledge & reasoning | 67.5 |
| Math | 60.8 |
| Instruction following | 51.3 |
| Code | 75.0 |
| German | 74.5 |
Tasks
| Category | Task | Score | ± | n |
|---|
| Knowledge & reasoning |
mmlu_pro |
79.3 |
± 2.3 |
300 |
| Knowledge & reasoning |
gpqa_diamond_gen |
63.1 |
± 3.4 |
198 |
| Knowledge & reasoning |
musr_gen |
60.0 |
± 3.1 |
250 |
| Math |
omni_math_gen |
60.8 |
± 3.1 |
250 |
| Instruction following |
ifbench |
38.3 |
± 2.8 |
300 |
| Instruction following |
ifeval |
90.7 |
± 2.4 |
150 |
| Code |
livecodebench_gen |
75.0 |
± 3.1 |
200 |
| German |
mmlu_prox_de |
78.3 |
± 2.4 |
300 |
| German |
include_de_gen |
66.9 |
± 3.8 |
154 |
± is the standard error; the 95% interval spans roughly 1.96 standard errors around the score.
Speed
| Measure | Context (target) | Context (actual) | Median | p10 | p90 | n |
|---|
| Generation [t/s] |
– |
– |
16.5 |
16.5 |
16.5 |
5 |
| Prompt processing [t/s] |
512 |
529 |
851.2 |
772.9 |
852.9 |
5 |
| Prompt processing [t/s] |
4096 |
4012 |
1359.0 |
1357.9 |
1445.3 |
5 |
| Prompt processing [t/s] |
16384 |
15971 |
1408.5 |
1406.8 |
1429.6 |
5 |
| Time to first token [ms] |
512 |
529 |
621.5 |
620.2 |
687.9 |
5 |
| Time to first token [ms] |
4096 |
4012 |
2952.1 |
2782.7 |
2954.6 |
5 |
| Time to first token [ms] |
16384 |
15971 |
11338.7 |
11172.4 |
11352.7 |
5 |
Measured directly at the inference server.