← Overview
qwen3.8-flash-next-abliterated (pool)
- Quantization
- W4A16-NVFP4
- Reasoning
- off
- Profile
- quick
- Suite
- v2b
- Date
- 2026-09-27
- Hardware
- 2× NVIDIA DGX Spark, je 128 GB RAM
- Inference software
- vLLM 0.26.0
- Measured
- directly at the inference server
- Overall
- 72.5
Categories
| Category | Score |
|---|
| Knowledge & reasoning | 77.2 |
| Math | 72.0 |
| Instruction following | 55.2 |
| Code | 77.0 |
| German | 81.4 |
Tasks
| Category | Task | Score | ± | n |
|---|
| Knowledge & reasoning |
mmlu_pro |
87.0 |
± 1.9 |
300 |
| Knowledge & reasoning |
gpqa_diamond_gen |
75.3 |
± 3.1 |
198 |
| Knowledge & reasoning |
musr_gen |
69.2 |
± 2.9 |
250 |
| Math |
omni_math_gen |
72.0 |
± 2.8 |
250 |
| Instruction following |
ifbench |
43.7 |
± 2.9 |
300 |
| Instruction following |
ifeval |
90.0 |
± 2.5 |
150 |
| Code |
livecodebench_gen |
77.0 |
± 3.0 |
200 |
| German |
mmlu_prox_de |
86.7 |
± 2.0 |
300 |
| German |
include_de_gen |
70.8 |
± 3.7 |
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] |
– |
– |
41.2 |
38.2 |
41.6 |
5 |
| Prompt processing [t/s] |
512 |
529 |
1429.4 |
878.3 |
1446.7 |
5 |
| Prompt processing [t/s] |
4096 |
4024 |
2569.1 |
2017.1 |
2570.7 |
5 |
| Prompt processing [t/s] |
16384 |
16001 |
2721.6 |
2264.6 |
2727.5 |
5 |
| Time to first token [ms] |
512 |
529 |
370.1 |
365.7 |
764.0 |
5 |
| Time to first token [ms] |
4096 |
4024 |
1566.3 |
1565.3 |
2090.1 |
5 |
| Time to first token [ms] |
16384 |
16001 |
5879.2 |
5866.5 |
7247.8 |
5 |
Measured directly at the inference server.