Benchmark score vs normalized inference time
Quality vs estimated response time on normalized hardware.
How to read estimates: normalized inference time is estimated execution time at 1 TB/s memory bandwidth, so model variants can be compared independently of a specific GPU. Estimated generation is throughput on normalized hardware, not a measured GPU benchmark.
Benchmark results analysis
Lower X and higher Y is better.
| Model | Tokens | Effective workload | Estimated generation | Benchmark score |
|---|---|---|---|---|
| Qwen 3.8 27B NVFP4 Think: medium 7 benchmarks • 8854 tests | IN48 342 439 OUT18 884 828 | 3225 tokens 98 s/task | 33.0 tok/s | 89.3% |
| Qwen 3.6 27B NVFP4 Thinking 7 benchmarks • 8854 tests | IN57 263 699 OUT21 305 322 | 3700 tokens 112 s/task | 33.0 tok/s | 87.7% |
| Qwen 3.8 Flash Next NVFP4 Instruct 6 benchmarks • 8422 tests | IN30 123 778 OUT3 938 488 | 1183 tokens 12 s/task | 99.0 tok/s | 80.2% |
| Qwen 3.6 27B NVFP4 Instruct 7 benchmarks • 8854 tests | IN81 807 475 OUT5 537 402 | 2473 tokens 75 s/task | 33.0 tok/s | 78.0% |
| Qwen 3.6 35B NVFP4 Thinking 7 benchmarks • 8854 tests | IN59 285 184 OUT27 730 792 | 4471 tokens 26 s/task | 172.0 tok/s | 76.6% |
| Qwen 3.8 27B NVFP4 Instruct 7 benchmarks • 8854 tests | IN75 160 602 OUT5 259 349 | 2292 tokens 69 s/task | 33.0 tok/s | 74.0% |
| Qwen 3.6 35B NVFP4 Instruct 7 benchmarks • 8854 tests | IN86 878 664 OUT13 576 249 | 3496 tokens 20 s/task | 172.0 tok/s | 55.1% |
| Qwen 3.8 Flash Next NVFP4 Think: medium 2 benchmarks • 493 tests | IN9 937 345 OUT1 945 238 | 7977 tokens 81 s/task | 99.0 tok/s | 33.8% |