Qwen3 235B-A22B: how much VRAM it needs and which GPU runs it

Qwen3 235B-A22B in Q4_K_M with an 8k context takes 134.9 GB of GPU memory. The cheapest option right now is 2× A800 PCIe 80 GB, from $1.98/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	Fits on
BF16	438.0	440.5	444.9	2+ GPUs
FP8	222.6	225.1	229.5	B300 288 GB
Q8_0	232.8	235.3	239.7	B300 288 GB
Q6_K	179.8	182.3	186.7	B300 288 GB
Q5_K_M	155.4	157.9	162.3	B200 180 GB
Q4_K_M	132.4	134.9	139.3	H200 NVL 141 GB
Q3_K_M	104.7	107.2	111.6	H200 NVL 141 GB
Q2_K	79.8	82.3	86.7	H100 NVL 94 GB

Context cache

How much memory the cache takes per request in FP16. A Q8 cache is about half. 188 KB per token.

8K context: 1.5 GB
32K context: 5.9 GB

Cheapest GPUs that run it

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

2× A800 PCIe 80 GB87% usedfrom $1.98/hRent
2× RTX PRO 5000 72 GB97% usedfrom $2.25/hRent
2× A100 PCIe 80 GB87% usedfrom $2.38/hRent
8× RTX 3090 24 GB76% usedfrom $2.41/hRent
8× RTX 4090 24 GB76% usedfrom $2.80/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 Qwen3 235B-A22B need?: In Q4_K_M with an 8k context, 134.9 GB: weights 132.4 GB, context cache 1.5 GB and about 1 GB of runtime memory. In BF16, 440.5 GB.
Does Qwen3 235B-A22B run on an RTX 4090 with 24 GB?: It does not fit on one RTX 4090 even in Q4_K_M: it needs 134.9 GB. A 141 GB GB card or several RTX 4090s will do.
How much does it cost to run Qwen3 235B-A22B?: The cheapest fitting server right now is 2× A800 PCIe 80 GB, from $1.98/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	Fits on
BF16	438.0	440.5	444.9	2+ GPUs
FP8	222.6	225.1	229.5	B300 288 GB
Q8_0	232.8	235.3	239.7	B300 288 GB
Q6_K	179.8	182.3	186.7	B300 288 GB
Q5_K_M	155.4	157.9	162.3	B200 180 GB
Q4_K_M	132.4	134.9	139.3	H200 NVL 141 GB
Q3_K_M	104.7	107.2	111.6	H200 NVL 141 GB
Q2_K	79.8	82.3	86.7	H100 NVL 94 GB

Cheapest GPUs that run it

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

2× A800 PCIe 80 GB87% usedfrom $1.98/hRent

2× RTX PRO 5000 72 GB97% usedfrom $2.25/hRent

2× A100 PCIe 80 GB87% usedfrom $2.38/hRent

8× RTX 3090 24 GB76% usedfrom $2.41/hRent

8× RTX 4090 24 GB76% usedfrom $2.80/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 Qwen3 235B-A22B need?

In Q4_K_M with an 8k context, 134.9 GB: weights 132.4 GB, context cache 1.5 GB and about 1 GB of runtime memory. In BF16, 440.5 GB.

Does Qwen3 235B-A22B run on an RTX 4090 with 24 GB?

It does not fit on one RTX 4090 even in Q4_K_M: it needs 134.9 GB. A 141 GB GB card or several RTX 4090s will do.

How much does it cost to run Qwen3 235B-A22B?

The cheapest fitting server right now is 2× A800 PCIe 80 GB, from $1.98/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.