Hunyuan Hy3: how much VRAM it needs and which GPU runs it

Hunyuan Hy3 in FP8 with an 8k context takes 282.8 GB of GPU memory. The cheapest option right now is 4× A100 PCIe 80 GB, from $4.76/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
BF16	≈556.5	560.0	567.5	597.5	2+ GPUs
FP8	279.3	282.8	290.3	320.3	2+ GPUs

Context cache

How much memory the cache takes per request in FP16. A Q8 cache is about half. 320 KB per token.

8K context: 2.5 GB
32K context: 10.0 GB
128K context: 40.0 GB

Cheapest GPUs that run it

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

4× A100 PCIe 80 GB92% usedfrom $4.76/hRent
4× A100 SXM4 80 GB92% usedfrom $6.04/hRent
8× RTX 4090 48 GB77% usedfrom $6.51/hRent
8× A100 PCIe 40 GB93% usedfrom $7.53/hRent
4× RTX PRO 6000 Max-Q 96 GB76% usedfrom $8.79/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 Hunyuan Hy3 need?: In FP8 with an 8k context, 282.8 GB: weights 279.3 GB, context cache 2.5 GB and about 1 GB of runtime memory. In BF16, 560.0 GB.
Does Hunyuan Hy3 run on an RTX 4090 with 24 GB?: It does not fit on one RTX 4090 even in FP8: it needs 282.8 GB. A 80+ GB card or several RTX 4090s will do.
How much does it cost to run Hunyuan Hy3?: The cheapest fitting server right now is 4× A100 PCIe 80 GB, from $4.76/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.

Quantisation

File

8K context

32K context

128K context

Fits on

BF16

≈556.5

560.0

567.5

597.5

2+ GPUs

FP8

279.3

282.8

290.3

320.3

2+ GPUs

Cheapest GPUs that run it

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

4× A100 PCIe 80 GB92% usedfrom $4.76/hRent

4× A100 SXM4 80 GB92% usedfrom $6.04/hRent

8× RTX 4090 48 GB77% usedfrom $6.51/hRent

8× A100 PCIe 40 GB93% usedfrom $7.53/hRent

4× RTX PRO 6000 Max-Q 96 GB76% usedfrom $8.79/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 Hunyuan Hy3 need?

In FP8 with an 8k context, 282.8 GB: weights 279.3 GB, context cache 2.5 GB and about 1 GB of runtime memory. In BF16, 560.0 GB.

Does Hunyuan Hy3 run on an RTX 4090 with 24 GB?

It does not fit on one RTX 4090 even in FP8: it needs 282.8 GB. A 80+ GB card or several RTX 4090s will do.

How much does it cost to run Hunyuan Hy3?

The cheapest fitting server right now is 4× A100 PCIe 80 GB, from $4.76/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.