Individual experiment result
MMLU-ProX (DE)
Kolibri-1 (FP8) Aleph Alpha
Exact match reasoning off fp8
67.0%
Exact match
Run details
- Date
- 2026-10-05
- Test cases
- 11,759
- Median latency
- 3.60s
- Total cost
- —
- Avg. prompt tokens
- 1,496.3
- Avg. answer tokens
- 942.8
- Avg. reasoning tokens
- —
- Quantization
- fp8
About this benchmark
German MMLU-ProX across 14 subjects; primary metric accuracy. Legacy rows have 11,759 items. Corrected v1.1 rows have 11,737 after a frozen, model-independent manifest excludes 20 duplicated-gold translation collisions, one independently reviewed ambiguous source item, and one independently confirmed notation-translation collapse; the row's item count identifies the definition.
Subject breakdown
| Subject | Score |
|---|---|
| biology | 82.1% |
| business | 70.0% |
| chemistry | 82.0% |
| computer science | 74.9% |
| economics | 71.2% |
| engineering | 53.1% |
| health | 61.1% |
| history | 53.8% |
| law | 36.2% |
| math | 86.0% |
| other | 53.8% |
| philosophy | 56.1% |
| physics | 71.3% |
| psychology | 69.0% |
Provenance
- Canonical model
- aleph-alpha/kolibri-1-fp8
- Run ID
2026-10-05T21-12-35__openai__Aleph-Alpha__Kolibri-1__mmlu_prox_de