Qwen 2.5 32B Instruct
Qwen 2.5 32B Instruct needs roughly 16.3GB VRAM at Q4 quantization (65.0GB at FP16). 46 GPUs we track can run it fully in VRAM at 8k context.
Alibaba32.5B params125k contextApache 2.0Commercial use ok
VRAM at each quantization
Assumes 8k context. KV cache grows linearly with context length.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP16 | 65.0 GB | 2.15 GB | 75.2 GB |
| Q8 | 32.5 GB | 2.15 GB | 38.8 GB |
| Q6_K | 24.4 GB | 2.15 GB | 29.7 GB |
| Q5_K_M | 20.3 GB | 2.15 GB | 25.2 GB |
| Q4_K_M | 16.3 GB | 2.15 GB | 20.6 GB |
| Q3_K_M | 13.0 GB | 2.15 GB | 17.0 GB |
| Q2_K | 9.8 GB | 2.15 GB | 13.3 GB |
Benchmarks
GPUs that run Qwen 2.5 32B Instruct natively (46)
- NVIDIA RTX 5090Q6_K · 73.5 t/s
- NVIDIA RTX 4090Q4_K_M · 62 t/s
- NVIDIA RTX 4080Q2_K · 73.5 t/s
- NVIDIA RTX 4060 Ti 16GBQ2_K · 29.5 t/s
- NVIDIA RTX 3090Q4_K_M · 57.6 t/s
- NVIDIA RTX 3090 TiQ4_K_M · 62 t/s
- NVIDIA H100 80GBFP16 · 51.5 t/s
- NVIDIA A100 80GBFP16 · 31.4 t/s
- NVIDIA A100 40GBQ6_K · 63.8 t/s
- NVIDIA L40SQ8 · 26.6 t/s
- NVIDIA RTX A6000Q8 · 23.6 t/s
- NVIDIA RTX 6000 AdaQ8 · 29.5 t/s
- AMD Radeon RX 7900 XTXQ4_K_M · 59.1 t/s
- AMD Radeon RX 7900 XTQ3_K_M · 61.5 t/s
- AMD Radeon RX 6800 XTQ2_K · 52.5 t/s
- AMD Instinct MI300XFP16 · 81.5 t/s
- Apple M4 Ultra (384GB)FP16 · 16.8 t/s
- Apple M4 Ultra (192GB)FP16 · 16.8 t/s
- Apple M4 Max (128GB)FP16 · 8.4 t/s
- Apple M4 Max (96GB)FP16 · 8.4 t/s
- Apple M4 Max (64GB)Q8 · 16.8 t/s
- Apple M4 Max (48GB)Q8 · 16.8 t/s
- Apple M4 Pro (48GB)Q8 · 8.4 t/s
- Apple M4 Pro (24GB)Q3_K_M · 21 t/s
- Apple M4 (32GB)Q5_K_M · 5.9 t/s
- Apple M3 Max (128GB)FP16 · 6.2 t/s
- Apple M3 Max (96GB)FP16 · 6.2 t/s
- Apple M3 Max (64GB)Q8 · 12.3 t/s
- Apple M3 Max (48GB)Q8 · 12.3 t/s
- Apple M3 Max (36GB)Q6_K · 16.4 t/s
- Apple M3 Pro (36GB)Q6_K · 6.2 t/s
- Apple M3 Pro (18GB)Q2_K · 15.4 t/s
- Apple M3 (24GB)Q3_K_M · 7.7 t/s
- Apple M2 Ultra (384GB)FP16 · 12.3 t/s
- Apple M2 Ultra (192GB)FP16 · 12.3 t/s
- Apple M2 Max (96GB)FP16 · 6.2 t/s
- Apple M2 Max (64GB)Q8 · 12.3 t/s
- Apple M2 Max (32GB)Q5_K_M · 19.7 t/s
- Apple M2 Pro (32GB)Q5_K_M · 9.8 t/s
- Apple M2 (24GB)Q3_K_M · 7.7 t/s
- Apple M1 Ultra (128GB)FP16 · 12.3 t/s
- Apple M1 Ultra (64GB)Q8 · 24.6 t/s
- Apple M1 Max (64GB)Q8 · 12.3 t/s
- Apple M1 Max (32GB)Q5_K_M · 19.7 t/s
- Apple M1 Pro (32GB)Q5_K_M · 9.8 t/s
- Intel Arc A770 16GBQ2_K · 57.4 t/s
Plus 6 GPUs that run it with CPU offload (slower)
- NVIDIA RTX 4070 TiQ6_K · 5.2 t/s
- NVIDIA RTX 4070Q6_K · 5.2 t/s
- NVIDIA RTX 4060Q6_K · 2.8 t/s
- NVIDIA RTX 3080 10GBQ6_K · 7.8 t/s
- NVIDIA RTX 3060 12GBQ6_K · 3.7 t/s
- CPU only (system RAM)Q5_K_M · 0.5 t/s
Notes
Sweet spot for a single 24GB GPU at Q4.