← Interactive LLM VRAM Calculator — try your own context length & quantization
How much VRAM to run Qwen2.5-Coder 32B? (FP16/Q8/Q6/Q4 requirements)
Architecture: 64 layers, GQA with 8 KV heads (same base architecture as Qwen2.5 32B). Figures are approximations for weights + an 8k-token FP16 KV cache + ~1 GB runtime overhead, based on typical published GGUF sizes. Actual usage varies by runtime and settings.
| Quantization | Weights | Total VRAM (8k context) | Fits on (example) |
|---|---|---|---|
| FP16 full precision (16-bit) | ~65 GB | ~68 GB | A100 / H100 (80 GB) |
| Q8_0 8-bit | ~34 GB | ~38 GB | 2×RTX 3090 or RTX A6000 (48 GB) |
| Q6_K ~6.6-bit | ~27 GB | ~30 GB | RTX 5090 (32 GB) |
| Q5_K_M ~5.7-bit | ~23 GB | ~26 GB | RTX 5090 (32 GB) |
| Q4_K_M ~4.8-bit | ~20 GB | ~23 GB | RTX 3090 / RTX 4090 (24 GB) |
How much VRAM do I need to run Qwen2.5-Coder 32B?
You need roughly ~23 GB of VRAM to run Qwen2.5-Coder 32B at Q4_K_M with an 8k-token context — that fits on RTX 3090 / RTX 4090 (24 GB). At Q8_0 plan for about ~38 GB, and full FP16 needs around ~68 GB. Qwen2.5-Coder 32B has the same memory footprint as Qwen2.5 32B: Q4_K_M fits (tightly) on a single 24 GB RTX 3090/4090 with 8k context, which is why it became a favorite local coding model. For longer coding contexts (32k+), budget several extra GB of KV cache or use KV quantization.
What affects these numbers?
Three things: the quantization (bytes per parameter), the KV cache (grows linearly with context length — the table assumes 8k tokens), and runtime overhead. Longer contexts need more memory; quantizing the KV cache (e.g. q8_0 cache in llama.cpp) roughly halves that part. Use the interactive calculator to plug in your exact context length and quantization.