Zero-shot Question Answering and Commonsense Reasoning on (ARC-C, HellaSwag, PIQA, WinoGrande)
55.8ARC-C AccuracyFP16
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| FP16Model=Qwen3-8B, System=TURBOQUANT, Throughput (tok/s)=37.72026.04 | 55.8 | 57.1 | 76.9 | 67.6 | 64.4 | |
| RATEQUANTBit-rate=4.0 bits, Model=Qwen3-8B, System=TURBOQUANT, Throughput (tok/s)=38.02026.04 | 54.1 | 56.6 | 76.4 | 69.5 | 64.2 | |
| UniformBit-rate=4.0 bits, Model=Qwen3-8B, System=TURBOQUANT, Throughput (tok/s)=38.12026.04 | 52.5 | 55.2 | 74.4 | 66.9 | 62.2 |