← Overview
ternary-bonsai-2-27b
- Quantization
- FP8
- Reasoning
- off
- Profile
- quick
- Suite
- v2b
- Date
- 2026-09-21 – 2026-09-23
- Hardware
- NVIDIA DGX Spark, 128 GB RAM
- Inference software
- vLLM 0.26.0
- Measured
- directly at the inference server
- Overall
- 61.3
Categories
| Category | Score |
|---|
| Knowledge & reasoning | 69.5 |
| Math | 59.6 |
| Instruction following | 46.3 |
| Code | 56.5 |
| German | 74.5 |
Tasks
| Category | Task | Score | ± | n |
|---|
| Knowledge & reasoning |
mmlu_pro |
77.7 |
± 2.4 |
300 |
| Knowledge & reasoning |
gpqa_diamond_gen |
65.7 |
± 3.4 |
198 |
| Knowledge & reasoning |
musr_gen |
65.2 |
± 3.0 |
250 |
| Math |
omni_math_gen |
59.6 |
± 3.1 |
250 |
| Instruction following |
ifbench |
34.3 |
± 2.7 |
300 |
| Instruction following |
ifeval |
82.7 |
± 3.1 |
150 |
| Code |
livecodebench_gen |
56.5 |
± 3.5 |
200 |
| German |
mmlu_prox_de |
77.3 |
± 2.4 |
300 |
| German |
include_de_gen |
68.8 |
± 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] |
– |
– |
28.5 |
28.5 |
28.5 |
5 |
| Prompt processing [t/s] |
512 |
529 |
622.5 |
602.7 |
638.3 |
5 |
| Prompt processing [t/s] |
4096 |
4024 |
898.1 |
896.5 |
899.6 |
5 |
| Prompt processing [t/s] |
16384 |
16001 |
892.4 |
886.4 |
903.4 |
5 |
| Time to first token [ms] |
512 |
529 |
849.8 |
828.8 |
876.9 |
5 |
| Time to first token [ms] |
4096 |
4024 |
4480.4 |
4473.0 |
4488.6 |
5 |
| Time to first token [ms] |
16384 |
16001 |
17931.1 |
17712.7 |
18051.2 |
5 |
Measured directly at the inference server.