Mathematical Reasoning on GSM8K lm-eval-harness 0.4.11
94.41GSM8K Flexible Extraction AccuracyXFP
Evaluation results
| Method | Links | ||
|---|---|---|---|
| XFPModel=Qwen3.5-122B-A10B, Hardware=RTX PRO 6000 Blackwell, Tensor Parallelism=2, Quantization=XFP, quantize-on-load, Effective bits=~3.97, VRAM / GPU=57 / 96 GiB, GPU util. (single-stream)=98% / 98%, GPU power (single-stream)=240 / 240 W2026.05 | 94.41 | 94.49 |