Qwen3 32B VRAM requirements

Qwen3 32B has 32.76 billion parameters in its published weight files. At 4 bits, the weights alone need approximately 16.4 GB. The default inference scenario estimates 20.9 GB including cache and runtime headroom.

Updated 2026-09-27. Memory estimates use decimal GB. Quantized weights may add scales, metadata and buffers; a compatible implementation and model licence are still required.

Memory at different precisions

How the estimate is calculated

Weights = parameters × bits ÷ 8. For supported standard attention, KV cache = 2 × layers × KV heads × head dimension × context tokens × batch size × 2 bytes. Inference adds the chosen percentage of runtime headroom to weights and cache. Divide bytes by 1,000,000,000 for GB.

The LoRA scenario keeps the chosen base weights and allows 18 bytes per trainable adapter parameter, plus the activation allowance. Full fine-tuning uses 18 bytes per parameter for a mixed-precision AdamW scenario, plus that allowance. Its total does not fall when you select 4-bit inference weights. These are budget estimates, not measured memory peaks; training context, optimizer, checkpointing and implementation can change them substantially.

Single-GPU candidates for this inference scenario

The offers update with the memory inputs above. Availability not reported means stock has not been confirmed. Memory capacity alone does not establish runtime compatibility or speed.

Estimate the rental cost of a training run · Compare an API bill with self-hosting

Other models from this publisher

• Qwen2.5 0.5B
• Qwen2.5 1.5B
• Qwen2.5 3B
• Qwen2.5 7B
• Qwen2.5 14B
• Qwen2.5 32B
• Qwen2.5 72B
• Qwen2.5 Coder 0.5B
• Qwen2.5 Coder 1.5B
• Qwen2.5 Coder 3B
• Qwen2.5 Coder 7B
• Qwen2.5 Coder 14B
• Qwen2.5 Coder 32B
• Qwen3 0.6B
• Qwen3 1.7B
• Qwen3 4B
• Qwen3 8B
• Qwen3 14B
• Qwen3 30B A3B
• Qwen3 235B A22B