← Overview
gemma-4-12b-it-heretic
- Quantization
- NVFP4
- Reasoning
- off
- Profile
- quick
- Suite
- v2b
- Date
- 2026-09-18
- Hardware
- NVIDIA DGX Spark, 128 GB RAM
- Inference software
- vLLM 0.26.0
- Measured
- directly at the inference server
- Overall
- 57.7
Categories
| Category | Score |
|---|
| Knowledge & reasoning | 63.7 |
| Math | 48.4 |
| Instruction following | 46.8 |
| Code | 62.0 |
| German | 67.7 |
Tasks
| Category | Task | Score | ± | n |
|---|
| Knowledge & reasoning |
mmlu_pro |
76.7 |
± 2.4 |
300 |
| Knowledge & reasoning |
gpqa_diamond_gen |
56.1 |
± 3.5 |
198 |
| Knowledge & reasoning |
musr_gen |
58.4 |
± 3.1 |
250 |
| Math |
omni_math_gen |
48.4 |
± 3.2 |
250 |
| Instruction following |
ifbench |
33.0 |
± 2.7 |
300 |
| Instruction following |
ifeval |
88.7 |
± 2.6 |
150 |
| Code |
livecodebench_gen |
62.0 |
± 3.4 |
200 |
| German |
mmlu_prox_de |
70.0 |
± 2.7 |
300 |
| German |
include_de_gen |
63.0 |
± 3.9 |
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] |
– |
– |
21.4 |
21.4 |
21.4 |
5 |
| Prompt processing [t/s] |
512 |
528 |
5070.5 |
4947.7 |
5118.3 |
5 |
| Prompt processing [t/s] |
4096 |
4014 |
5258.0 |
5221.7 |
5293.9 |
5 |
| Prompt processing [t/s] |
16384 |
15985 |
3514.1 |
3505.2 |
3518.5 |
5 |
| Time to first token [ms] |
512 |
528 |
103.9 |
103.2 |
106.7 |
5 |
| Time to first token [ms] |
4096 |
4014 |
763.4 |
758.2 |
768.7 |
5 |
| Time to first token [ms] |
16384 |
15985 |
4548.8 |
4543.1 |
4560.4 |
5 |
Measured directly at the inference server.