← Overview
qwen3.8-27b-abliterated
- Quantization
- NVFP4
- Reasoning
- off
- Profile
- quick
- Suite
- v2b
- Date
- 2026-09-20 – 2026-09-21
- Hardware
- NVIDIA DGX Spark, 128 GB RAM
- Inference software
- vLLM 0.26.0
- Measured
- directly at the inference server
- Overall
- 61.4
Categories
| Category | Score |
|---|
| Knowledge & reasoning | 71.5 |
| Math | 58.0 |
| Instruction following | 46.1 |
| Code | 59.5 |
| German | 72.1 |
Tasks
| Category | Task | Score | ± | n |
|---|
| Knowledge & reasoning |
mmlu_pro |
81.0 |
± 2.3 |
300 |
| Knowledge & reasoning |
gpqa_diamond_gen |
70.7 |
± 3.2 |
198 |
| Knowledge & reasoning |
musr_gen |
62.8 |
± 3.1 |
250 |
| Math |
omni_math_gen |
58.0 |
± 3.1 |
250 |
| Instruction following |
ifbench |
34.0 |
± 2.7 |
300 |
| Instruction following |
ifeval |
82.7 |
± 3.1 |
150 |
| Code |
livecodebench_gen |
59.5 |
± 3.5 |
200 |
| German |
mmlu_prox_de |
74.3 |
± 2.5 |
300 |
| German |
include_de_gen |
67.5 |
± 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] |
– |
– |
15.9 |
15.1 |
16.0 |
5 |
| Prompt processing [t/s] |
512 |
529 |
1805.5 |
1685.0 |
1818.2 |
5 |
| Prompt processing [t/s] |
4096 |
4024 |
2472.0 |
2175.2 |
2483.2 |
5 |
| Prompt processing [t/s] |
16384 |
16001 |
1685.4 |
1489.9 |
1688.7 |
5 |
| Time to first token [ms] |
512 |
529 |
293.0 |
291.0 |
314.3 |
5 |
| Time to first token [ms] |
4096 |
4024 |
1627.8 |
1620.5 |
1872.0 |
5 |
| Time to first token [ms] |
16384 |
16001 |
9493.9 |
9475.4 |
10858.2 |
5 |
Measured directly at the inference server.