Written by Jakub Rusinowski · Last updated April 28, 2025
Yes, but it is tight
Yes, but it is tight — Qwen 3 32B at Q4_K_M needs about 22.8 GB of the 24 GB usable on RTX 4090 Desktop (24 GB VRAM, 64 GB RAM), leaving only ~1.2 GB before the runtime starts swapping. Expect ~33.5 tok/s (estimated), with room for about 8,192 tokens of context.
Confidence: medium · Recommended quantization: Q4_K_M · Estimated speed: ~33.5 tok/s
| Usable memory for models | 24 GB |
| Memory bandwidth | 1008 GB/s |
| Form factor | Desktop |
| Operating system | Windows or Linux |
| Memory upgradeable | Yes |
| Quant | Memory needed | Fits 24 GB? | Max context | Est. speed | Download |
|---|---|---|---|---|---|
| F16 | 68.5 GB | ✗ No | — | — | 65.6 GB |
| Q8_0 | 37.8 GB | ✗ No | — | — | 34.8 GB |
| Q6_K | 29.8 GB | ✗ No | — | — | 26.9 GB |
| Q5_K_M | 26.2 GB | ✗ No | — | — | 23.2 GB |
| Q4_K_M | 22.8 GB | ✓ Yes | 8K | ~33.5 tok/s | 19.8 GB |
| Q3_K_M | 16.9 GB | ✓ Yes | 32K | ~45.1 tok/s | 14 GB |
| Q2_K | 13.7 GB | ✓ Yes | 32K | ~55.8 tok/s | 10.8 GB |
| Variant | Needs at Q4_K_M | Fits? | Est. speed |
|---|---|---|---|
| Qwen 3 235B-A22B (MoE) | 144.3 GB | ✗ Too large | — |
| Qwen 3 32B | 22.8 GB | ✓ Fits | ~33.5 tok/s |
| Qwen 3 30B-A3B (MoE) | 20 GB | ✓ Fits | ~184.7 tok/s |
| Qwen 3 14B | 11.1 GB | ✓ Fits | ~66.9 tok/s |
| Qwen 3 8B | 7 GB | ✓ Fits | ~104.2 tok/s |
Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving
Yes, but it is tight — Qwen 3 32B at Q4_K_M needs about 22.8 GB of the 24 GB usable on RTX 4090 Desktop (24 GB VRAM, 64 GB RAM), leaving only ~1.2 GB before the runtime starts swapping. Expect ~33.5 tok/s (estimated), with room for about 8,192 tokens of context.
Q4_K_M — it needs about 22.8 GB of the 24 GB available, downloads as roughly 19.8 GB, and runs at an estimated 33.5 tokens/sec with up to 8K of context.
Nothing binding — the model fits with headroom and generates at a usable speed on this hardware.
Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving