Qwen3 32B GPU requirements: 16.4 GB at 4-bit, 65.5 GB at BF16
Qwen3 32B needs 65.5 GB of GPU memory for its weights at BF16, 32.8 GB at FP8 and 16.4 GB at 4-bit, plus 8.59 GB of KV cache for a 32K context. The cheapest node that runs it here is 2× RTX A4000 at 4-bit, for $0.06 an hour on spot.
Weights plus a 32K 16-bit KV cache, with 10% of the node's memory kept free. Billed per minute.
- Parameters
- 32.8BDense
- Weights, 4-bit
- 16.4 GBBF16: 65.5 GB
- KV cache, 32K
- 8.59 GBOne sequence, 16-bit cache
- Context window
- 128K40K native, extended with YaRN
How much VRAM Qwen3 32B needs, format by format.
Weights are parameters × bytes per parameter. Each context column adds the KV cache of one sequence of that length.
| Format | Weights | + 8Kcontext | + 32Kcontext | + 128Kcontext | Cheapest node, 32K |
|---|---|---|---|---|---|
| BF16Published weights, fine-tuning | 65.5 GB | 67.7 GB | 74.1 GB | 99.9 GB | 2× A40 · $0.12/h |
| FP8Hopper, Ada, Blackwell, MI300 | 32.8 GB | 34.9 GB | 41.3 GB | 67.1 GB | 1× L40 · $0.09/h |
| 4-bitAWQ, GPTQ, GGUF Q4 | 16.4 GB | 18.5 GB | 25.0 GB | 50.7 GB | 2× RTX A4000 · $0.06/h |
The serving engine adds its own buffers, and an 8-bit KV cache halves the context columns. We keep 10% of a node's memory free when we say a model fits. Speed is a separate question: it follows memory bandwidth and, for a mixture of experts, the active parameters.
GPUs that run Qwen3 32B, cheapest first.
For each format, the smallest node of each GPU model that holds the weights and a 32K context, sorted by spot price. One row per GPU model; larger nodes of the same GPU work too.
| Format | GPU | GPU memory | Spot | On-demand | 730 h spot | Action |
|---|---|---|---|---|---|---|
| 4-bit | 2× RTX A4000Ampere · 25.0 GB needed | 32 GB | $0.06 | $0.10 | $43.80 | Deploy |
| 4-bit | 1× A40Ampere · 25.0 GB needed | 48 GB | $0.06 | $0.09 | $43.80 | Deploy |
| 4-bit | 2× T4Turing · 25.0 GB needed | 32 GB | $0.08 | $0.14 | $58.40 | Deploy |
| 4-bit | 1× RTX PRO 4500 BlackwellBlackwell · 25.0 GB needed | 32 GB | $0.09 | $0.19 | $65.70 | Deploy |
| FP8 | 1× L40Ada Lovelace · 41.3 GB needed | 48 GB | $0.09 | $0.19 | $65.70 | Deploy |
| FP8 | 1× RTX 6000 AdaAda Lovelace · 41.3 GB needed | 48 GB | $0.09 | $0.19 | $65.70 | Deploy |
| FP8 | 2× L4Ada Lovelace · 41.3 GB needed | 48 GB | $0.10 | $0.16 | $73.00 | Deploy |
| FP8 | 2× RTX 4090Ada Lovelace · 41.3 GB needed | 48 GB | $0.12 | $0.18 | $87.60 | Deploy |
| BF16 | 2× A40Ampere · 74.1 GB needed | 96 GB | $0.12 | $0.18 | $87.60 | Deploy |
| BF16 | 2× L40Ada Lovelace · 74.1 GB needed | 96 GB | $0.18 | $0.38 | $131.40 | Deploy |
| BF16 | 2× RTX 6000 AdaAda Lovelace · 74.1 GB needed | 96 GB | $0.18 | $0.38 | $131.40 | Deploy |
| BF16 | 2× RTX A6000Ampere · 74.1 GB needed | 96 GB | $0.18 | $0.34 | $131.40 | Deploy |
Spot capacity can be reclaimed with a 2-minute notice; for a live endpoint, use on-demand. How spot works · serving a language model.
What a long context costs in memory.
Grouped-query attention, 8 KV heads of 128 across 64 layers.
| Context | KV cache, 1 sequence | 8 sequences |
|---|---|---|
| 8K tokens | 2.15 GB | 17.2 GB |
| 32K tokens | 8.59 GB | 68.7 GB |
| 128K tokens | 34.4 GB | 275 GB |
Serving several users
KV cacheEvery concurrent request keeps its own cache, so memory for context grows with users × context length. A server batching 8 requests of 32K needs eight times the single-sequence figure — 68.7 GB here.
- Recommended node2× RTX A4000
- Weights at 4-bit16.4 GB
- Context that fits, 1 sequence46K
Paged attention (vLLM, SGLang) allocates the cache in blocks as requests grow, so a node holds more short requests than the worst case suggests.
Qwen3 32B, as published.
- Developer
- Alibaba QwenDense reasoning model, thinking and non-thinking modes
- Parameters
- 32.8B
- Architecture
- Dense model, 64 layersGrouped-query attention, 8 KV heads of 128 across 64 layers
- Context window
- 131,072 tokens40,960 native, the rest with YaRN scaling
- Published format
- BF16, 65.5 GB
- Licence
- Apache 2.0
- Released
- 29 April 2025
- Source
- Qwen/Qwen3-32Bconfig.json read 26 October 2026
Serve it with vLLM
template: vLLMOn 2× RTX A4000, from the instance terminal:
$ vllm serve Qwen/Qwen3-32B \ --tensor-parallel-size 2 \ --max-model-len 32768
Very recent architectures need a recent vLLM; the model card gives the minimum version and any extra flags.
Qwen3 32B on a GPU — the questions that come up
How much VRAM does Qwen3 32B need?
For the weights alone: 65.5 GB at BF16, 32.8 GB at FP8 and 16.4 GB at 4-bit. Add 8.59 GB of KV cache for one 32K-token context, or 34.4 GB at the full 128K window, and keep about 10% of the memory free for activations.
Can Qwen3 32B run on a single GPU?
One RTX 4090 (24 GB): no. One RTX 5090 (32 GB): yes, at 4-bit. One H100 SXM (80 GB): yes, at FP8. One H200 SXM (141 GB): yes, at BF16. One B200 SXM (180 GB): yes, at BF16. Those answers assume a 32K context and the widest format that still fits.
What is the cheapest way to run Qwen3 32B?
Here, 2× RTX A4000 at 4-bit: $0.06 per hour on spot and $0.10 on-demand, billed per minute — $43.80 for a 730-hour month on spot. That node has 32 GB of GPU memory for the 25.0 GB this setup needs.
How long a context fits on the recommended node?
With the weights at 4-bit on 2× RTX A4000, the remaining memory holds a 16-bit KV cache for about 47,378 tokens of context for one sequence (46K). An 8-bit KV cache doubles that; serving several users at once divides it between them.
Which format should I run it in?
BF16 keeps the published weights untouched and is the format for fine-tuning. FP8 halves the memory and runs in hardware on Hopper, Ada, Blackwell and MI300-class GPUs. 4-bit quantisation (AWQ, GPTQ, GGUF) quarters it, at a quality cost that is larger on small models than on big ones; check it against your own evaluation before relying on it.
How do I deploy it?
Pick the node in the table above and launch it with the vLLM template; the model downloads from Hugging Face (Qwen/Qwen3-32B) on first start. Spread the weights over the node with --tensor-parallel-size equal to the GPU count, and set --max-model-len to the context you need. New architectures need a recent vLLM: check the model card for the minimum version.
Sizing another model? GPU requirements for 64 open models · serving a language model · fine-tuning.
Qwen3 32B on 2× RTX A4000 for $0.06 an hour.
Launch the node with the vLLM template, point it at Qwen/Qwen3-32B, stop it when you are done. Prepaid, billed by the minute.
Billed per minute from the moment the instance is reachable. Minimum credit $30, no subscription.