Qwen3 32B
26 records across 9 hardware.
| hardware | model | quant | backend | tok/s | pp tok/s | W | ttft s | ctx/tg | date | class | flags | source |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RTX 5090 | Qwen3 32B 32B | Q4_K | llama.cpp | 61.4 | 2931.3 | – | – | 4096 | – | sourced | – | source |
| RTX 5090 | Qwen3 32B 32B | Q4_K | llama.cpp | 50.9 | 2077.2 | – | – | 16384 | – | sourced | – | source |
| RTX Pro 5000 Blackwell | Qwen3 32B 32B | Q4_K | llama.cpp | 46.6 | 2030.3 | – | – | 4096 | – | sourced | – | source |
| RTX 5090 | Qwen3 32B 32B | Q4_K | llama.cpp | 43.8 | 1451.1 | – | – | 32768 | – | sourced | – | source |
| RTX Pro 5000 Blackwell | Qwen3 32B 32B | Q4_K | llama.cpp | 42.0 | 1497.7 | – | – | 16384 | – | sourced | – | source |
| RTX 4090 | Qwen3 32B 32B | Q4_K | llama.cpp | 39.6 | 2448.0 | – | – | 4096 | – | sourced | – | source |
| RTX 3090 Ti | Qwen3 32B 32B | Q4_K | llama.cpp | 38.0 | 1238.8 | – | – | 4096 | – | sourced | – | source |
| RTX PRO 4500 Blackwell | Qwen3 32B | Q4_K - Medium | llama.cpp (CUDA) | 37.93 | 1675.78 | – | – | tg128 | 2025-11-14 | community | – | source |
| RTX 6000 Ada | Qwen3 32B 32B | Q4_K | llama.cpp | 37.4 | 1203.9 | – | – | 4096 | – | sourced | – | source |
| RTX Pro 5000 Blackwell | Qwen3 32B 32B | Q4_K | llama.cpp | 36.4 | 1079.8 | – | – | 32768 | – | sourced | – | source |
| RTX PRO 4500 Blackwell | Qwen3 32B | Q4_K - Medium | llama.cpp (CUDA) | 36.38 | 1514.35 | – | – | tg128 | 2025-11-14 | community | – | source |
| RTX 4090 | Qwen3 32B 32B | Q4_K | llama.cpp | 34.4 | 1707.4 | – | – | 16384 | – | sourced | – | source |
| RTX 3090 Ti | Qwen3 32B 32B | Q4_K | llama.cpp | 32.6 | 862.7 | – | – | 16384 | – | sourced | – | source |
| V100 | Qwen3 32B | Q4_K - Medium | llama.cpp (CUDA) | 31.8 | 613.13 | – | – | tg128 | 2025-09-22 | community | – | source |
| V100 | Qwen3 32B | Q4_K - Medium | llama.cpp (CUDA) | 30.74 | 642.41 | – | – | tg128 | 2025-09-22 | community | – | source |
| RTX Pro 5000 Blackwell | Qwen3 32B 32B | Q4_K | llama.cpp | 28.9 | 629.6 | – | – | 65536 | – | sourced | – | source |
| RTX 6000 Ada | Qwen3 32B 32B | Q4_K | llama.cpp | 25.1 | 977.5 | – | – | 16384 | – | sourced | – | source |
| RTX A6000 | Qwen3 32B 32B | Q4_K | llama.cpp | 24.6 | 736.2 | – | – | 4096 | – | sourced | – | source |
| RTX A6000 | Qwen3 32B 32B | Q4_K | llama.cpp | 18.3 | 704.8 | – | – | 16384 | – | sourced | – | source |
| RTX 6000 Ada | Qwen3 32B 32B | Q4_K | llama.cpp | 18.2 | 595.5 | – | – | 32768 | – | sourced | – | source |
| M5 Pro | Qwen3 32B | 4bit | oMLX | 14.6 | 323.0 | – | – | 1024 | 2026-10-14 | community | – | source |
| RTX A6000 | Qwen3 32B 32B | Q4_K | llama.cpp | 13.5 | 502.7 | – | – | 32768 | – | sourced | – | source |
| M5 Pro | Qwen3 32B | 4bit | oMLX | 13.5 | 325.9 | – | – | 4096 | 2026-10-14 | community | – | source |
| M5 Pro | Qwen3 32B | 4bit | oMLX | 12.7 | 301.8 | – | – | 8192 | 2026-10-14 | community | – | source |
| RTX 6000 Ada | Qwen3 32B 32B | Q4_K | llama.cpp | 11.8 | 348.6 | – | – | 65536 | – | sourced | – | source |
| RTX A6000 | Qwen3 32B 32B | Q4_K | llama.cpp | 8.7 | 319.0 | – | – | 65536 | – | sourced | – | source |