Qwen3.5
Benchmark
The following three-machine comparison is based on the full formal matrix of lzc_thor running on 2026-03-20T05-47-33Z, retaining only the complete results from correct usage, without including experimental single-request data such as p1, cublas, or mmq.
| Model | Best Parallel | Orin | Thor T4000 | Thor T5000 |
|---|---|---|---|---|
Qwen3.5-0.8B-Q4_K_M | 16 | 1326.86 tok/s / TTFT 1.94 s / 42.54 W | 1311.50 tok/s / TTFT 1.88 s / 21.72 W | 1783.26 tok/s / TTFT 1.37 s / 30.35 W |
Qwen3.5-2B-Q4_K_M | 16 | 989.57 tok/s / TTFT 2.91 s / 46.43 W | 960.84 tok/s / TTFT 3.01 s / 22.46 W | 1291.47 tok/s / TTFT 2.17 s / 32.93 W |
Qwen3.5-4B-Q4_K_M | 16 | 431.96 tok/s / TTFT 7.16 s / 48.48 W | 442.65 tok/s / TTFT 7.07 s / 23.95 W | 622.09 tok/s / TTFT 4.87 s / 35.87 W |
Qwen3.5-9B-Q4_K_M | 16 | 318.65 tok/s / TTFT 10.85 s / 56.49 W | 314.55 tok/s / TTFT 11.23 s / 24.60 W | 450.86 tok/s / TTFT 7.71 s / 35.85 W |
Qwen3.5-27B-Q4_K_M | 16 | 95.80 tok/s / TTFT 44.38 s / 56.82 W | 90.82 tok/s / TTFT 48.44 s / 25.32 W | 150.03 tok/s / TTFT 25.49 s / 36.91 W |
Qwen3.5-35B-A3B-Q4_K_M | 16 | 192.59 tok/s / TTFT 27.41 s / 43.12 W | 181.57 tok/s / TTFT 30.34 s / 18.74 W | 271.35 tok/s / TTFT 19.09 s / 31.05 W |
Qwen3.5-122B-A10B-IQ2_S | 16 | 65.14 tok/s / TTFT 46.02 s / 42.48 W | 60.92 tok/s / TTFT 50.95 s / 19.49 W | 76.55 tok/s / TTFT 43.03 s / 26.25 W |
Latest T4000 SM110 Retest
2026-03-20T17-10-00Z's latest sm110 maxn retest has now completed Thor T4000, and is therefore listed separately to avoid mixing in incomplete T5000 / Orin data into the same comparison table.
| Model | Best Parallel | Thor T4000 |
|---|---|---|
Qwen3.5-0.8B-Q4_K_M | 16 | 1313.47 tok/s / TTFT 1.88 s / 21.73 W |
Qwen3.5-2B-Q4_K_M | 16 | 960.84 tok/s / TTFT 3.01 s / 22.84 W |
Qwen3.5-4B-Q4_K_M | 16 | 442.74 tok/s / TTFT 7.07 s / 23.95 W |
Qwen3.5-9B-Q4_K_M | 16 | 314.46 tok/s / TTFT 11.23 s / 24.60 W |
Qwen3.5-27B-Q4_K_M | 16 | 91.00 tok/s / TTFT 48.40 s / 25.33 W |
Qwen3.5-35B-A3B-Q4_K_M | 16 | 181.36 tok/s / TTFT 30.35 s / 18.74 W |
Qwen3.5-122B-A10B-IQ2_S | 16 | 60.67 tok/s / TTFT 51.23 s / 19.49 W |
Notes
- Comparison data from three machines:
/home/catdog/lzc_thor/docs/runs/2026-03-20T05-47-33Z - Source of the latest
sm110retest:/home/catdog/lzc_thor/docs/runs/2026-03-20T17-10-00Z Qwen3.5-397B-A17B-IQ1_Sdid not form valid benchmarks withT4000andOrinin the most recent test, thus it is not included in the official benchmark table.- Power consumption column records the maximum power observed during testing.