DeepSeek V3.2 GPU requirements: VRAM and cheapest GPU
DeepSeek V3.2 has 685.4B parameters, published in FP8. It is a mixture-of-experts model: 8 of its 256 experts work on each token, about 51.9B parameters, so it generates at roughly the speed of a model that size. Every expert still has to sit in GPU memory, which is why the memory figures below are for the full model. It is released under mit, which allows commercial use.
- Parameters
- 685.4B
- Smallest footprint (INT4)
- about 397 GB
- Cheapest GPU per hour (FP16)
- $8.99
- Type
- Text
deepseek-ai/DeepSeek-V3.2
- Parameters 685.4B
- Active per token 51.9B
- Type Text
- Task text-generation
- Max context 160K
- Licence mit
- Downloads 8.4M
- Likes 1.6K
Calculating for DeepSeek V3.2
More settings
DeepSeek V3.2 needs about 1,509 GB of GPU memory
Inference (vLLM), FP16 / BF16, 8K tokens context, one request at a time.
Where the memory goes
- Weights 1,371 GB
- KV cache 0.58 GB
- Runtime overhead 138 GB
- Total 1,509 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 | 6 × B300 288GBCheapestBlackwell Ultra · HBM3e | $8.99 | 6sold as 8 GPUs; you need 6 | $71.92 | $52,502 | RunPod (opens in a new tab)8x B300 |
| More than 8: needs several nodes: MI300X 192GB, GB200 186GB, B200 180GB, H200 141GB, GH200 96GB, H100 NVL 94GB, RTX PRO 6000 96GB, H100 80GB, A100 80GB, L40S 48GB, RTX 6000 Ada 48GB, RTX A6000 48GB, A40 48GB, A100 40GB, RTX 5090 32GB, RTX 4090 24GB, RTX 3090 24GB, L4 24GB, A10 24GB, T4 16GB, V100 16GB | ||||||
| Provider | Instance | Per GPU / hr | GPUs needed | Total / hr | Total / month (24/7) |
|---|---|---|---|---|---|
| RunPod (opens in a new tab)Cheapest | 8x B300us | $8.99 | 6 × B300 288GB sold as 8 GPUs; you need 6 | about $71.92 | about $52,502 |
| Nebius (opens in a new tab) | 8x B300 288GBeu/us | $9.50 | 6 × B300 288GB sold as 8 GPUs; you need 6 | about $76.00 | about $55,480 |
| AWS (opens in a new tab) | p6-b300.48xlargeus-east-1 | $17.80 | 6 × B300 288GB sold as 8 GPUs; you need 6 | about $142.42 | about $103,964 |
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 | 1,509 GB | 1,760 GB | – | 12,073 GB |
| INT8 | 793 GB | – | – | – |
| INT4 | 397 GB | – | 648 GB | – |
Cheapest GPU by precision
Inference with an 8K context and one request at a time.
FP16 / BF16
about 1,509 GB · 8K tokens
The weights as published. Nothing lost, but the largest footprint.
6 × B300 288GB
RunPod (opens in a new tab) · 8x B300
sold as 8 GPUs; you need 6- Per GPU / hr
- $8.99
- GPUs needed
- 6
- Total / hr
- about $71.92
- per month, 24/7, approx.
- $52,502
INT8
about 793 GB · 8K tokens
Weights at half the size. Most chat and RAG work shows no visible difference.
5 × MI300X 192GB
RunPod (opens in a new tab) · 8x MI300X
sold as 8 GPUs; you need 5- Per GPU / hr
- $2.39
- GPUs needed
- 5
- Total / hr
- about $19.12
- per month, 24/7, approx.
- $13,958
INT4
about 397 GB · 8K tokens
Weights at a quarter of the size. Small drops show up first in maths and code.
3 × MI300X 192GB
RunPod (opens in a new tab) · 4x MI300X
sold as 4 GPUs; you need 3- Per GPU / hr
- $2.39
- GPUs needed
- 3
- Total / hr
- about $9.56
- per month, 24/7, approx.
- $6,979
Inference vs fine-tuning
Inference (FP16)
1×
of the memory inference needs
1,509 GB
6 × B300 288GB · about $71.92
LoRA fine-tune
1.2×
of the memory inference needs
1,760 GB
7 × B300 288GB · about $71.92
QLoRA fine-tune
0.4×
of the memory inference needs
648 GB
4 × MI300X 192GB · about $9.56
Full fine-tune
8×
of the memory inference needs
12,073 GB
more than one 8-GPU node
Serving it at FP16 fits on 6 × B300 288GB (RunPod). LoRA needs 7 × B300 288GB (RunPod), QLoRA 4 × MI300X 192GB (RunPod) and a full fine-tune more than one 8-GPU node. 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 397 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
Not realistically. Even at INT4 it needs about 397 GB, more than the largest Mac can give its GPU.
Similar models
Open models of a similar size.
- DeepSeek R1Text684.5B · about 1,508 GB FP16 · 1.7M ↓GPU needs
- DeepSeek V3Text684.5B · about 1,508 GB FP16 · 1.4M ↓GPU needs
- K-EXAONE 2.0Text749.4B · about 1,650 GB FP16 · 14.3K ↓GPU needs
- GLM-5.3Text753.3B · about 1,659 GB FP16 · 1.9M ↓GPU needs
- GLM 5Text753.9B · about 1,660 GB FP16 · 1.2M ↓GPU needs
- GLM 5.1Text753.9B · about 1,660 GB FP16 · 789.8K ↓GPU needs
DeepSeek V3.2: common questions
How much VRAM does DeepSeek V3.2 need?
About 1,509 GB at FP16, 793 GB at INT8 and 397 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 DeepSeek V3.2?
6 × B300 288GB (about $8.99 per GPU / hour; sold as 8 GPUs; you need 6) on RunPod, at about $71.92 an hour for FP16 / BF16 inference with an 8K context. Prices collected Nov 13, 2026; they change often.
Can I fine-tune DeepSeek V3.2 on a single GPU?
Not on one GPU. Even QLoRA needs about 648 GB, more than an 80 GB GPU holds. LoRA needs about 1,760 GB and a full fine-tune about 12,073 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 DeepSeek V3.2 run on a consumer GPU like the RTX 4090?
Not on its own. Even at INT4 it needs about 397 GB, more than the card’s 24 GB. Two or more cards, or a bigger data-centre GPU, will do it.
Does DeepSeek V3.2 fit on a single H100 80GB?
No. Even at INT4 it needs about 397 GB, so you need several GPUs.
How many GPUs do I need for DeepSeek V3.2?
At FP16 with an 8K context: 21 × 80 GB, 35 × 48 GB or 70 × 24 GB GPUs. At INT4 you need about a quarter of the memory for the weights.
How much VRAM does DeepSeek V3.2 need at INT8?
About 793 GB: 720 GB for the weights, 0.58 GB for an 8K KV cache and 73 GB of runtime overhead.
How much does it cost to run DeepSeek V3.2 24/7 for a month?
About $52,502 on the cheapest setup we found (6 × B300 288GB (about $8.99 per GPU / hour; sold as 8 GPUs; you need 6), $71.92 an hour for 730 hours). The cheapest on AWS, Google Cloud or Azure is about $103,964 a month.
FP16 vs INT8 vs INT4: which should I use?
FP16 keeps full quality and needs the most memory (about 1,509 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 397 GB in total, with a small drop in quality that shows most in maths and code.
Can DeepSeek V3.2 run on a Mac?
Not realistically. Even at INT4 it needs about 397 GB, more than the largest Mac can give its 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.