Qwen3 32B GPU requirements: VRAM and cheapest GPU
Qwen3 32B has 32.8B parameters, published in BF16. It is released under apache-2.0, which allows commercial use.
- Parameters
- 32.8B
- Smallest footprint (INT4)
- about 22 GB
- Cheapest GPU per hour (FP16)
- $0.50
- Type
- Text
Qwen/Qwen3-32B
- Parameters 32.8B
- Type Text
- Task text-generation
- Max context 40K
- Licence apache-2.0
- Downloads 3.1M
- Likes 794
Calculating for Qwen3 32B
More settings
Qwen3 32B needs about 75 GB of GPU memory
Inference (vLLM), FP16 / BF16, 8K tokens context, one request at a time.
Where the memory goes
- Weights 66 GB
- KV cache 2.1 GB
- Runtime overhead 7.8 GB
- Total 75 GB
GPU counts keep 10% of each card free for spikes. Plan extra if you expect long prompts or many users at once.
| Rank | Configuration | Per GPU / hr | GPUs needed | Total / hr | Total / month (24/7) | Provider |
|---|---|---|---|---|---|---|
| 1 | 1 × RTX PRO 6000 96GBCheapestBlackwell · GDDR7 | $0.50 | 1 | $0.50 | $365 | RunPod (opens in a new tab)1x RTX PRO 6000 MaxQ |
| 2 | 4 × RTX 3090 24GBAmpere · GDDR6X | $0.26 | 4 | $1.06 | $772 | Vast.ai (opens in a new tab)4x RTX 3090 (median of 5 offers) |
| 3 | 2 × A40 48GBAmpere · GDDR6 | $0.59 | 2 | $1.18 | $861 | RunPod (opens in a new tab)2x A40 |
| 4 | 2 × RTX A6000 48GBAmpere · GDDR6 | $0.69 | 2 | $1.38 | $1,007 | RunPod (opens in a new tab)2x RTX A6000 |
| 5 | 2 × L40S 48GBAda Lovelace · GDDR6 | $0.79 | 2 | $1.57 | $1,150 | Vast.ai (opens in a new tab)2x L40S (median of 4 offers) |
| 6 | 1 × GH200 96GBHopper · HBM3 | $1.99 | 1 | $1.99 | $1,453 | Vultr (opens in a new tab)vbm-72c-480gb-gh200-gpu (bare metal) |
| 7 | 4 × RTX 4090 24GBAda Lovelace · GDDR6X | $0.53 | 4 | $2.14 | $1,561 | Vast.ai (opens in a new tab)4x RTX 4090 (median of 14 offers) |
| 8 | 2 × RTX 6000 Ada 48GBAda Lovelace · GDDR6 | $1.09 | 2 | $2.18 | $1,591 | RunPod (opens in a new tab)2x RTX 6000 Ada |
| 9 | 1 × MI300X 192GBCDNA 3 · HBM3 | $2.39 | 1 | $2.39 | $1,745 | RunPod (opens in a new tab)1x MI300X |
| 10 | 4 × L4 24GBAda Lovelace · GDDR6 | $0.69 | 4 | $2.76 | $2,015 | RunPod (opens in a new tab)4x L4 |
| 11 | 3 × RTX 5090 32GBBlackwell · GDDR7 | $0.73 | 3sold as 4 GPUs; you need 3 | $2.94 | $2,143 | Vast.ai (opens in a new tab)4x RTX 5090 (median of 35 offers) |
| 12 | 1 × H100 NVL 94GBHopper · HBM3 | $3.19 | 1 | $3.19 | $2,329 | RunPod (opens in a new tab)1x H100 NVL |
| 13 | 1 × H200 141GBHopper · HBM3e | $3.79 | 1 | $3.79 | $2,767 | RunPod (opens in a new tab)1x H200 NVL |
| 14 | 2 × A100 80GBAmpere · HBM2e | $2.09 | 2 | $4.18 | $3,051 | RunPod (opens in a new tab)2x A100 PCIe |
| 15 | 4 × A10 24GBAmpere · GDDR6 | $1.42 | 4 | $5.67 | $4,141 | AWS (opens in a new tab)g5.12xlarge |
| 16 | 2 × H100 80GBHopper · HBM3 | $2.89 | 2 | $5.78 | $4,219 | RunPod (opens in a new tab)2x H100 PCIe |
| 17 | 6 × V100 16GBVolta · HBM2 | $0.79 | 6sold as 8 GPUs; you need 6 | $6.32 | $4,614 | Lambda (opens in a new tab)8x Tesla V100 |
| 18 | 6 × T4 16GBTuring · GDDR6 | $0.98 | 6sold as 8 GPUs; you need 6 | $7.82 | $5,712 | AWS (opens in a new tab)g4dn.metal |
| 19 | 3 × A100 40GBAmpere · HBM2 | $1.99 | 3sold as 4 GPUs; you need 3 | $7.96 | $5,811 | Lambda (opens in a new tab)4x A100 PCIe |
| 20 | 1 × B200 180GBBlackwell · HBM3e | $7.99 | 1 | $7.99 | $5,833 | RunPod (opens in a new tab)1x B200 |
| 21 | 1 × B300 288GBBlackwell Ultra · HBM3e | $8.99 | 1 | $8.99 | $6,563 | RunPod (opens in a new tab)1x B300 |
| 22 | 1 × GB200 186GBBlackwell · HBM3e | $10.50 | 1sold as 4 GPUs; you need 1 | $42.00 | $30,660 | CoreWeave (opens in a new tab)GB200 NVL72 |
| Provider | Instance | Per GPU / hr | GPUs needed | Total / hr | Total / month (24/7) |
|---|---|---|---|---|---|
| RunPod (opens in a new tab)Cheapest | 1x RTX PRO 6000 MaxQus | $0.50 | 1 × RTX PRO 6000 96GB | about $0.50 | about $365 |
| Vast.ai (opens in a new tab) | 4x RTX 3090 (median of 5 offers)worldwide marketplace | $0.26 | 4 × RTX 3090 24GB | about $1.06 | about $772 |
| Nebius (opens in a new tab) | 1x RTX PRO 6000 96GBeu/us | $1.80 | 1 × RTX PRO 6000 96GB | about $1.80 | about $1,314 |
| Vultr (opens in a new tab) | vbm-72c-480gb-gh200-gpu (bare metal)us | $1.99 | 1 × GH200 96GB | about $1.99 | about $1,453 |
| Lambda (opens in a new tab) | 2x A6000us | $1.09 | 2 × RTX A6000 48GB | about $2.18 | about $1,591 |
| AWS (opens in a new tab) | g7e.2xlargeus-east-1 | $3.36 | 1 × RTX PRO 6000 96GB | about $3.36 | about $2,455 |
| Google Cloud (opens in a new tab) | g2-standard-48us-central1 | $1.00 | 4 × L4 24GB | about $4.00 | about $2,921 |
| CoreWeave (opens in a new tab) | GH200us | $6.50 | 1 × GH200 96GB | about $6.50 | about $4,745 |
| Azure (opens in a new tab) | NC40ads_H100_v5eastus | $6.98 | 1 × H100 NVL 94GB | about $6.98 | about $5,095 |
Prices collected Nov 13, 2026. Every cost is approximate: on-demand, US dollars, mostly US regions (marketplaces such as Vast.ai are worldwide).
Memory by precision and workload
| Precision | Inference | LoRA fine-tune | QLoRA fine-tune | Full fine-tune |
|---|---|---|---|---|
| FP16 / BF16 | 75 GB | 92 GB | – | 585 GB |
| INT8 | 41 GB | – | – | – |
| INT4 | 22 GB | – | 39 GB | – |
Cheapest GPU by precision
Inference with an 8K context and one request at a time.
FP16 / BF16
about 75 GB · 8K tokens
The weights as published. Nothing lost, but the largest footprint.
1 × RTX PRO 6000 96GB
RunPod (opens in a new tab) · 1x RTX PRO 6000 MaxQ
- Per GPU / hr
- $0.50
- GPUs needed
- 1
- Total / hr
- about $0.50
- per month, 24/7, approx.
- $365
INT8
about 41 GB · 8K tokens
Weights at half the size. Most chat and RAG work shows no visible difference.
1 × RTX PRO 6000 96GB
RunPod (opens in a new tab) · 1x RTX PRO 6000 MaxQ
- Per GPU / hr
- $0.50
- GPUs needed
- 1
- Total / hr
- about $0.50
- per month, 24/7, approx.
- $365
INT4
about 22 GB · 8K tokens
Weights at a quarter of the size. Small drops show up first in maths and code.
2 × V100 16GB
Vast.ai (opens in a new tab) · 2x Tesla V100 (median of 4 offers)
- Per GPU / hr
- $0.15
- GPUs needed
- 2
- Total / hr
- about $0.30
- per month, 24/7, approx.
- $220
Inference vs fine-tuning
Inference (FP16)
1×
of the memory inference needs
75 GB
1 × RTX PRO 6000 96GB · about $0.50
LoRA fine-tune
1.2×
of the memory inference needs
92 GB
3 × A40 48GB · about $2.36
QLoRA fine-tune
0.5×
of the memory inference needs
39 GB
1 × RTX PRO 6000 96GB · about $0.50
Full fine-tune
7.8×
of the memory inference needs
585 GB
4 × MI300X 192GB · about $9.56
Serving it at FP16 fits on 1 × RTX PRO 6000 96GB (RunPod). LoRA needs 3 × A40 48GB (RunPod), QLoRA 1 × RTX PRO 6000 96GB (RunPod) and a full fine-tune 4 × MI300X 192GB (RunPod). Training keeps activations and optimiser state in memory as well as the weights, which is why it so often needs more GPUs than serving the same model.
Will it run on a Mac or a single RTX 4090?
RTX 4090 (24 GB)
Not on its own. Even at INT4 it needs about 22 GB, more than the card’s 24 GB. Two or more cards, or a bigger data-centre GPU, will do it.
Mac with Apple silicon
Yes, with enough memory. At INT4 it needs about 22 GB, so a Mac with 32 GB or more of unified memory (macOS lets the GPU use about three quarters of it). Expect slower generation than on a data-centre GPU.
Similar models
Open models of a similar size.
- Qwen2.5 32B InstructText32.8B · about 75 GB FP16 · 739.2K ↓GPU needs
- Qwen2.5 Coder 32B InstructText32.8B · about 75 GB FP16 · 544.4K ↓GPU needs
- Gemma 4 31B StyleTuneText32.7B · about 74 GB FP16 · 820 ↓GPU needs
- Nanbeige4.3-33BText33.1B · about 76 GB FP16 · 22.3K ↓GPU needs
- Laguna XS.2Text33.4B · about 75 GB FP16 · 764.4K ↓GPU needs
- Laguna XS 2.1Text33.4B · about 75 GB FP16 · 615.8K ↓GPU needs
Qwen3 32B: common questions
How much VRAM does Qwen3 32B need?
About 75 GB at FP16, 41 GB at INT8 and 22 GB at INT4, with an 8K context and one request at a time. Longer contexts and more parallel requests add KV cache on top.
What is the cheapest GPU to run Qwen3 32B?
1 × RTX PRO 6000 96GB (about $0.50 per GPU / hour) on RunPod, at about $0.50 an hour for FP16 / BF16 inference with an 8K context. Prices collected Nov 13, 2026; they change often.
Can I fine-tune Qwen3 32B on a single GPU?
Yes, with QLoRA: it needs about 39 GB, which fits one 80 GB GPU. LoRA needs about 92 GB and a full fine-tune about 585 GB.
How are these numbers calculated?
Weights are parameters times bytes per parameter (2 at FP16, 1 at INT8, 0.5 at INT4). The KV cache comes from the model’s layers, KV heads, head size and context length. We add about 10% plus 1 GB for the runtime and fill each GPU to 90%. Prices are on-demand list prices in US regions, collected from each provider’s public pricing.
Can Qwen3 32B run on a consumer GPU like the RTX 4090?
Not on its own. Even at INT4 it needs about 22 GB, more than the card’s 24 GB. Two or more cards, or a bigger data-centre GPU, will do it.
Does Qwen3 32B fit on a single H100 80GB?
Yes, at INT8, INT4 with an 8K context.
How many GPUs do I need for Qwen3 32B?
At FP16 with an 8K context: 2 × 80 GB, 2 × 48 GB or 4 × 24 GB GPUs. At INT4 you need about a quarter of the memory for the weights.
How much VRAM does Qwen3 32B need at INT8?
About 41 GB: 34 GB for the weights, 2.1 GB for an 8K KV cache and 4.7 GB of runtime overhead.
How much does it cost to run Qwen3 32B 24/7 for a month?
About $365 on the cheapest setup we found (1 × RTX PRO 6000 96GB (about $0.50 per GPU / hour), $0.50 an hour for 730 hours). The cheapest on AWS, Google Cloud or Azure is about $2,455 a month.
FP16 vs INT8 vs INT4: which should I use?
FP16 keeps full quality and needs the most memory (about 75 GB here). INT8 halves the weights with little loss for most chat work. INT4 (AWQ, GPTQ or GGUF Q4) cuts them to a quarter, about 22 GB in total, with a small drop in quality that shows most in maths and code.
Can Qwen3 32B run on a Mac?
Yes, with enough memory. At INT4 it needs about 22 GB, so a Mac with 32 GB or more of unified memory (macOS lets the GPU use about three quarters of it). Expect slower generation than on a data-centre GPU.
vLLM or llama.cpp?
vLLM is built for serving many requests at once on NVIDIA and AMD GPUs and reserves most of the GPU memory for KV cache. llama.cpp runs GGUF files on almost anything, including Macs and CPUs, and suits one user or a small team. Memory needs are similar for the same precision.
Prices collected Nov 13, 2026. Every cost is approximate: on-demand, US dollars, mostly US regions (marketplaces such as Vast.ai are worldwide).
How we calculate
- Weights. Parameters times bytes per parameter: 2 at FP16, 1 at INT8, 0.5 at INT4, plus about 5% for the quantisation scales at INT8 and INT4.
- KV cache. 2 × layers × KV heads × head size × context × requests × 2 bytes, from the model’s config.json. Where a setting is missing we estimate it and say so.
- Overhead. About 10% plus 1 GB for the runtime, in line with what vLLM reserves.
- Fine-tuning. LoRA keeps 16-bit weights and adds a small adapter, its optimiser and activations. QLoRA does the same over 4-bit weights. A full fine-tune needs about 16 bytes per parameter with mixed-precision Adam (weights, gradients and optimiser states), plus activations.
- GPUs. We fill each GPU to 90% of its memory and stop at 8 per node. Costs are on-demand list prices in US regions; marketplace clouds show the median offer.