Decoding Throughput on Ultrafeedback
88.8AR Throughput (tok/s)Autoregressive
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| AutoregressiveModel=Qwen-3 32B/0.6B2026.03 | 88.8 | — | — | — | — | |
| AutoregressiveModel=Llama-3.1-Instruct 70B/1B2026.03 | 54.7 | — | — | — | — | |
| Speculative DecodingModel=Llama-3.1-Instruct 70B/1B2026.03 | — | 138 | — | — | — | |
| Speculative DecodingModel=Qwen-3 32B/0.6B2026.03 | — | 122 | — | — | — | |
| SSDModel=Llama-3.1-Instruct 70B/1B2026.03 | — | — | 215 | 1.55 | 3.93 | |
| SSDModel=Qwen-3 32B/0.6B2026.03 | — | — | 174 | 1.43 | 1.96 |