Can RTX 4090 run Qwen2.5 32B Instruct?
Comfortable: EXL2 3.5bpw needs 16.0 GB at 4K, batch 1; headroom 8.0 GB.
Text input, FP16 KV cache, one card. Unified-memory capacity uses the GPU guide’s usable limit; negative headroom may require host offload. Speed source settings may differ from this 4K estimate.
Measured runs
| Framework | Quant | tok/s | VRAM used (GB) |
|---|---|---|---|
| llama.cpp | GGUF Q4_K_M | 44 | 22 |
Quant memory at 4K
| Quant | Total GB | Verdict | Weight sizing |
|---|---|---|---|
| GGUF Q3_K_M | 17.5 | Comfortable | Estimated weights |
| GGUF Q4_K_M | 21.7 | Tight | Estimated weights |
| EXL2 3.5bpw | 16.0 | Comfortable | Estimated weights |
Context capacity · EXL2 3.5bpw
| Context | Total GB | Verdict |
|---|---|---|
| 4K | 16.0 | Comfortable |
| 16K | 19.3 | Comfortable |
| 32K | 23.7 | Tight |
| 128K | 50.1 | Does not fit |
Generation speed
Measured: 44 tok/s, GGUF Q4_K_M, llama.cpp
EXL2 3.5bpw: 75 tok/s · Bandwidth ceiling (not measured)