Qwen3 VL 8B Instruct GPU requirements: VRAM and cheapest GPU
Qwen3 VL 8B Instruct has 8.8B parameters, published in BF16. It accepts up to 256K tokens of context. At that length the KV cache alone takes about 39 GB per request, so long prompts can change which GPU you need. It is released under apache-2.0, which allows commercial use.
- Parameters
- 8.8B
- Smallest footprint (INT4)
- about 7.4 GB
- Cheapest GPU per hour (FP16)
- $0.15
- Type
- Vision
Qwen/Qwen3-VL-8B-Instruct
- Parameters 8.8B
- Type Vision
- Task image-text-to-text
- Max context 256K
- Licence apache-2.0
- Downloads 7.1M
- Likes 1.3K
Calculating for Qwen3 VL 8B Instruct
More settings
Qwen3 VL 8B Instruct needs about 22 GB of GPU memory
Inference (vLLM), FP16 / BF16, 8K tokens context, one request at a time.
Where the memory goes
- Weights 18 GB
- KV cache 1.2 GB
- Runtime overhead 2.9 GB
- Total 22 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 | 2 × V100 16GBCheapestVolta · HBM2 | $0.15 | 2 | $0.30 | $220 | Vast.ai (opens in a new tab)2x Tesla V100 (median of 4 offers) |
| 2 | 1 × RTX PRO 6000 96GBBlackwell · GDDR7 | $0.50 | 1 | $0.50 | $365 | RunPod (opens in a new tab)1x RTX PRO 6000 MaxQ |
| 3 | 2 × RTX 3090 24GBAmpere · GDDR6X | $0.29 | 2 | $0.58 | $421 | Vast.ai (opens in a new tab)2x RTX 3090 (median of 18 offers) |
| 4 | 1 × A40 48GBAmpere · GDDR6 | $0.59 | 1 | $0.59 | $431 | RunPod (opens in a new tab)1x A40 |
| 5 | 1 × RTX 6000 Ada 48GBAda Lovelace · GDDR6 | $0.67 | 1 | $0.67 | $488 | Vast.ai (opens in a new tab)1x RTX 6000Ada (median of 8 offers) |
| 6 | 1 × A100 40GBAmpere · HBM2 | $0.68 | 1 | $0.68 | $497 | Vast.ai (opens in a new tab)1x A100 SXM4 (median of 8 offers) |
| 7 | 1 × RTX A6000 48GBAmpere · GDDR6 | $0.69 | 1 | $0.69 | $504 | RunPod (opens in a new tab)1x RTX A6000 |
| 8 | 1 × L40S 48GBAda Lovelace · GDDR6 | $0.78 | 1 | $0.78 | $571 | Vast.ai (opens in a new tab)1x L40S (median of 6 offers) |
| 9 | 2 × L4 24GBAda Lovelace · GDDR6 | $0.42 | 2 | $0.83 | $606 | Vast.ai (opens in a new tab)2x L4 (median of 4 offers) |
| 10 | 1 × RTX 5090 32GBBlackwell · GDDR7 | $0.87 | 1 | $0.87 | $634 | Vast.ai (opens in a new tab)1x RTX 5090 (median of 55 offers) |
| 11 | 2 × RTX 4090 24GBAda Lovelace · GDDR6X | $0.53 | 2 | $1.06 | $772 | Vast.ai (opens in a new tab)2x RTX 4090 (median of 24 offers) |
| 12 | 1 × A100 80GBAmpere · HBM2e | $1.21 | 1 | $1.21 | $884 | Vast.ai (opens in a new tab)1x A100 SXM4 (median of 8 offers) |
| 13 | 1 × GH200 96GBHopper · HBM3 | $1.99 | 1 | $1.99 | $1,453 | Vultr (opens in a new tab)vbm-72c-480gb-gh200-gpu (bare metal) |
| 14 | 1 × MI300X 192GBCDNA 3 · HBM3 | $2.39 | 1 | $2.39 | $1,745 | RunPod (opens in a new tab)1x MI300X |
| 15 | 1 × H100 80GBHopper · HBM3 | $2.89 | 1 | $2.89 | $2,110 | RunPod (opens in a new tab)1x H100 PCIe |
| 16 | 1 × H100 NVL 94GBHopper · HBM3 | $3.19 | 1 | $3.19 | $2,329 | RunPod (opens in a new tab)1x H100 NVL |
| 17 | 1 × H200 141GBHopper · HBM3e | $3.79 | 1 | $3.79 | $2,767 | RunPod (opens in a new tab)1x H200 NVL |
| 18 | 2 × T4 16GBTuring · GDDR6 | $0.98 | 2sold as 4 GPUs; you need 2 | $3.91 | $2,856 | AWS (opens in a new tab)g4dn.12xlarge |
| 19 | 2 × A10 24GBAmpere · GDDR6 | $1.42 | 2sold as 4 GPUs; you need 2 | $5.67 | $4,141 | AWS (opens in a new tab)g5.12xlarge |
| 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) |
|---|---|---|---|---|---|
| Vast.ai (opens in a new tab)Cheapest | 2x Tesla V100 (median of 4 offers)worldwide marketplace | $0.15 | 2 × V100 16GB | about $0.30 | about $220 |
| RunPod (opens in a new tab) | 1x RTX PRO 6000 MaxQus | $0.50 | 1 × RTX PRO 6000 96GB | about $0.50 | about $365 |
| Lambda (opens in a new tab) | 1x A6000us | $1.09 | 1 × RTX A6000 48GB | about $1.09 | about $796 |
| Nebius (opens in a new tab) | 1x L40S 48GBeu/us | $1.55 | 1 × L40S 48GB | about $1.55 | about $1,132 |
| Vultr (opens in a new tab) | vcg-l40s-16c-180g-48vramus | $1.67 | 1 × L40S 48GB | about $1.67 | about $1,220 |
| AWS (opens in a new tab) | g6e.xlargeus-east-1 | $1.86 | 1 × L40S 48GB | about $1.86 | about $1,359 |
| Google Cloud (opens in a new tab) | g2-standard-24us-central1 | $1.00 | 2 × L4 24GB | about $2.00 | about $1,461 |
| Azure (opens in a new tab) | NC24ads_A100_v4eastus | $3.67 | 1 × A100 80GB | about $3.67 | about $2,681 |
| CoreWeave (opens in a new tab) | GH200us | $6.50 | 1 × GH200 96GB | about $6.50 | about $4,745 |
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 | 22 GB | 27 GB | – | 159 GB |
| INT8 | 12 GB | – | – | – |
| INT4 | 7.4 GB | – | 13 GB | – |
Cheapest GPU by precision
Inference with an 8K context and one request at a time.
FP16 / BF16
about 22 GB · 8K tokens
The weights as published. Nothing lost, but the largest footprint.
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
INT8
about 12 GB · 8K tokens
Weights at half the size. Most chat and RAG work shows no visible difference.
1 × RTX 3090 24GB
Vast.ai (opens in a new tab) · 1x RTX 3090 (median of 29 offers)
- Per GPU / hr
- $0.25
- GPUs needed
- 1
- Total / hr
- about $0.25
- per month, 24/7, approx.
- $183
INT4
about 7.4 GB · 8K tokens
Weights at a quarter of the size. Small drops show up first in maths and code.
1 × RTX 3090 24GB
Vast.ai (opens in a new tab) · 1x RTX 3090 (median of 29 offers)
- Per GPU / hr
- $0.25
- GPUs needed
- 1
- Total / hr
- about $0.25
- per month, 24/7, approx.
- $183
Inference vs fine-tuning
Inference (FP16)
1×
of the memory inference needs
22 GB
2 × V100 16GB · about $0.30
LoRA fine-tune
1.3×
of the memory inference needs
27 GB
2 × V100 16GB · about $0.30
QLoRA fine-tune
0.6×
of the memory inference needs
13 GB
1 × RTX 3090 24GB · about $0.25
Full fine-tune
7.4×
of the memory inference needs
159 GB
4 × A40 48GB · about $2.36
Serving it at FP16 fits on 2 × V100 16GB (Vast.ai). LoRA needs 2 × V100 16GB (Vast.ai), QLoRA 1 × RTX 3090 24GB (Vast.ai) and a full fine-tune 4 × A40 48GB (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)
Yes, at INT8 with an 8K context: about 12 GB fits in the card’s 24 GB.
Mac with Apple silicon
Yes, with enough memory. At INT4 it needs about 7.4 GB, so a Mac with 16 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.
- Molmo2-8BVision8.7B · about 21 GB FP16 · 45.2K ↓GPU needs
- Qwen2.5 VL 7B InstructVision8.3B · about 20 GB FP16 · 5.8M ↓GPU needs
- Qwen2 VL 7B InstructVision8.3B · about 20 GB FP16 · 533.4K ↓GPU needs
- Moondream 3.1 (9B-A2B)Vision9.3B · about 23 GB FP16 · 9.6K ↓GPU needs
- MiniCPM-o 4.5Vision9.4B · about 23 GB FP16 · 179K ↓GPU needs
- Qwen3.5 9BVision9.7B · about 23 GB FP16 · 6.2M ↓GPU needs
Qwen3 VL 8B Instruct: common questions
How much VRAM does Qwen3 VL 8B Instruct need?
About 22 GB at FP16, 12 GB at INT8 and 7.4 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 VL 8B Instruct?
2 × V100 16GB (about $0.15 per GPU / hour) on Vast.ai, at about $0.30 an hour for FP16 / BF16 inference with an 8K context. Prices collected Nov 13, 2026; they change often.
Can I fine-tune Qwen3 VL 8B Instruct on a single GPU?
Yes, with QLoRA: it needs about 13 GB, which fits one 80 GB GPU. LoRA needs about 27 GB and a full fine-tune about 159 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 VL 8B Instruct run on a consumer GPU like the RTX 4090?
Yes, at INT8 with an 8K context: about 12 GB fits in the card’s 24 GB.
Does Qwen3 VL 8B Instruct fit on a single H100 80GB?
Yes, at FP16 / BF16, INT8, INT4 with an 8K context.
How many GPUs do I need for Qwen3 VL 8B Instruct?
At FP16 with an 8K context: 1 × 80 GB, 1 × 48 GB or 2 × 24 GB GPUs. At INT4 you need about a quarter of the memory for the weights.
How much VRAM does Qwen3 VL 8B Instruct need at INT8?
About 12 GB: 9.2 GB for the weights, 1.2 GB for an 8K KV cache and 2 GB of runtime overhead.
How much does it cost to run Qwen3 VL 8B Instruct 24/7 for a month?
About $220 on the cheapest setup we found (2 × V100 16GB (about $0.15 per GPU / hour), $0.30 an hour for 730 hours). The cheapest on AWS, Google Cloud or Azure is about $1,359 a month.
FP16 vs INT8 vs INT4: which should I use?
FP16 keeps full quality and needs the most memory (about 22 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 7.4 GB in total, with a small drop in quality that shows most in maths and code.
Can Qwen3 VL 8B Instruct run on a Mac?
Yes, with enough memory. At INT4 it needs about 7.4 GB, so a Mac with 16 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.