Qwen3 32B — VRAM requirements
Alibaba · 32.76B params · up to 32K context
| Quantization | VRAM at 4K context | VRAM at 32K context |
|---|---|---|
| FP16 / BF16 | 76.1 GB | 81.5 GB |
| FP8 | 38.4 GB | 43.8 GB |
| INT8 | 38.4 GB | 43.8 GB |
| INT4 | 19.6 GB | 25.0 GB |
Weights = params × bytes-per-parameter for the selected quantization.KV cache is always computed at FP16 and assumes batch size 1. A flat +15% is added on top for CUDA context/activation overhead — a rule of thumb, not a measured figure. Full methodology on the interactive calculator.
Compatible GPUs at 32K context
FP16 / BF16 (81.5 GB)
FP8 (43.8 GB)
INT8 (43.8 GB)
INT4 (25.0 GB)
paramsB is the exact safetensors total from the Hugging Face API for Qwen/Qwen3-32B (ungated); architecture fields from that repo's config.json. maxContextLength is Qwen's documented native 32,768 (131,072 with YaRN), not config.json's 40,960. Dense model — no MoE routing, so every parameter is active per token. Pulled 2026-09-19.