DeepSeek V3.2: how much VRAM it needs and which GPU runs it

DeepSeek V3.2 in Q4_K_M with an 8k context takes 379.1 GB of GPU memory. The cheapest option right now is 8× RTX PRO 5000 72 GB, from $9.03/h.

Find a server Open in the calculator

VRAM by quantisation

The weight file and the total GPU memory with the context cache. One GPU, one request.

Quantisation	File	8K context	32K context	128K context	Fits on
FP8native	642.1	643.7	645.3	651.7	2+ GPUs
BF16	1,250.1	1,251.6	1,253.2	1,259.7	2+ GPUs
Q8_0	664.3	665.8	667.4	673.9	2+ GPUs
Q6_K	513.5	515.0	516.6	523.0	2+ GPUs
Q5_K_M	443.5	445.0	446.6	453.0	2+ GPUs
Q4_K_M	377.6	379.1	380.7	387.1	2+ GPUs
Q3_K_M	298.2	299.8	301.4	307.8	2+ GPUs
Q2_K	228.5	230.0	231.7	238.1	B300 288 GB

Context cache

How much memory the cache takes per request in FP16. A Q8 cache is about half. 69 KB per token.

8K context: 0.54 GB
32K context: 2.1 GB
128K context: 8.6 GB

Cheapest GPUs that run it

GPUs from the catalog the model fits on, by price per hour.

8× RTX PRO 5000 72 GB69% usedfrom $9.03/hRent
8× RTX PRO 6000 96 GB52% usedfrom $18.25/hRent
8× RTX PRO 6000 Server 96 GB52% usedfrom $22.54/hRent
4× H200 NVL 141 GB69% usedfrom $27.04/hRent
4× H200 SXM 141 GB69% usedfrom $35.13/hRent

The price is the cheapest GPU of the model × the GPU count. A 2–8 GPU server can cost a little more per GPU.

How to run it on gpu.nz

For an API under load there is the vLLM template with an OpenAI-compatible API.

Questions

How much VRAM does DeepSeek V3.2 need?: In Q4_K_M with an 8k context, 379.1 GB: weights 377.6 GB, context cache 0.54 GB and about 1 GB of runtime memory. In BF16, 1,251.6 GB.
Does DeepSeek V3.2 run on an RTX 4090 with 24 GB?: It does not fit on one RTX 4090 even in Q4_K_M: it needs 379.1 GB. A 80+ GB card or several RTX 4090s will do.
How much does it cost to run DeepSeek V3.2?: The cheapest fitting server right now is 8× RTX PRO 5000 72 GB, from $9.03/h. Billing is per minute: stop the server and only its disk is billed.
Which quantisation should I pick?: Q4_K_M is the chat default: close to the original quality at a third of BF16's size. Q5_K_M and Q6_K are more accurate, Q8_0 is almost the original. For vLLM use FP8 or BF16.

File sizes checked against Hugging Face on 2026-10-19. 1 GB here is 1024³ bytes, like GPU memory. Hugging Face lists decimal gigabytes, so its numbers are 7% larger.

VRAM by quantisation

The weight file and the total GPU memory with the context cache. One GPU, one request.

Quantisation	File	8K context	32K context	128K context	Fits on
FP8native	642.1	643.7	645.3	651.7	2+ GPUs
BF16	1,250.1	1,251.6	1,253.2	1,259.7	2+ GPUs
Q8_0	664.3	665.8	667.4	673.9	2+ GPUs
Q6_K	513.5	515.0	516.6	523.0	2+ GPUs
Q5_K_M	443.5	445.0	446.6	453.0	2+ GPUs
Q4_K_M	377.6	379.1	380.7	387.1	2+ GPUs
Q3_K_M	298.2	299.8	301.4	307.8	2+ GPUs
Q2_K	228.5	230.0	231.7	238.1	B300 288 GB

Cheapest GPUs that run it

GPUs from the catalog the model fits on, by price per hour.

8× RTX PRO 5000 72 GB69% usedfrom $9.03/hRent

8× RTX PRO 6000 96 GB52% usedfrom $18.25/hRent

8× RTX PRO 6000 Server 96 GB52% usedfrom $22.54/hRent

4× H200 NVL 141 GB69% usedfrom $27.04/hRent

4× H200 SXM 141 GB69% usedfrom $35.13/hRent

The price is the cheapest GPU of the model × the GPU count. A 2–8 GPU server can cost a little more per GPU.

Questions

How much VRAM does DeepSeek V3.2 need?

In Q4_K_M with an 8k context, 379.1 GB: weights 377.6 GB, context cache 0.54 GB and about 1 GB of runtime memory. In BF16, 1,251.6 GB.

Does DeepSeek V3.2 run on an RTX 4090 with 24 GB?

It does not fit on one RTX 4090 even in Q4_K_M: it needs 379.1 GB. A 80+ GB card or several RTX 4090s will do.

How much does it cost to run DeepSeek V3.2?

The cheapest fitting server right now is 8× RTX PRO 5000 72 GB, from $9.03/h. Billing is per minute: stop the server and only its disk is billed.

Which quantisation should I pick?

Q4_K_M is the chat default: close to the original quality at a third of BF16's size. Q5_K_M and Q6_K are more accurate, Q8_0 is almost the original. For vLLM use FP8 or BF16.