Qwen3 14B: how much VRAM it needs and which GPU runs it

Qwen3 14B in Q4_K_M with an 8k context takes 10.7 GB of GPU memory. The cheapest option right now is RTX 3060 12 GB, from $0.084/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	27.5	29.8	33.6	Tesla V100 Super 32 GB
FP8	15.2	17.5	21.2	RTX 4000 Ada 20 GB
Q8_0	14.6	16.9	20.7	RTX 4000 Ada 20 GB
Q6_K	11.3	13.6	17.3	RTX 4070 Ti Super 16 GB
Q5_K_M	9.8	12.1	15.8	RTX 4070 Ti Super 16 GB
Q4_K_M	8.4	10.7	14.4	RTX 3060 12 GB
Q3_K_M	6.8	9.1	12.9	RTX 3080 10 GB
Q2_K	5.4	7.6	11.4	RTX 4060 8 GB

Context cache

How much memory the cache takes per request in FP16. A Q8 cache is about half. 160 KB per token.

8K context: 1.3 GB
32K context: 5.0 GB

Cheapest GPUs that run it

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

RTX 3060 12 GB91% usedfrom $0.084/hRent
RTX 4070 Super 12 GB91% usedfrom $0.193/hRent
RTX 4070 Ti Super 16 GB68% usedfrom $0.196/hRent
RTX A4000 16 GB68% usedfrom $0.204/hRent
RTX 4060 Ti 16 GB68% usedfrom $0.254/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

In the Ollama template the model is one command away:

ollama run qwen3:14b

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

Questions

How much VRAM does Qwen3 14B need?: In Q4_K_M with an 8k context, 10.7 GB: weights 8.4 GB, context cache 1.3 GB and about 1 GB of runtime memory. In BF16, 29.8 GB.
Does Qwen3 14B run on an RTX 4090 with 24 GB?: Yes. In Q4_K_M the model takes 10.7 GB and fits in 24 GB.
How much does it cost to run Qwen3 14B?: The cheapest fitting server right now is RTX 3060 12 GB, from $0.084/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	27.5	29.8	33.6	Tesla V100 Super 32 GB
FP8	15.2	17.5	21.2	RTX 4000 Ada 20 GB
Q8_0	14.6	16.9	20.7	RTX 4000 Ada 20 GB
Q6_K	11.3	13.6	17.3	RTX 4070 Ti Super 16 GB
Q5_K_M	9.8	12.1	15.8	RTX 4070 Ti Super 16 GB
Q4_K_M	8.4	10.7	14.4	RTX 3060 12 GB
Q3_K_M	6.8	9.1	12.9	RTX 3080 10 GB
Q2_K	5.4	7.6	11.4	RTX 4060 8 GB

Cheapest GPUs that run it

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

RTX 3060 12 GB91% usedfrom $0.084/hRent

RTX 4070 Super 12 GB91% usedfrom $0.193/hRent

RTX 4070 Ti Super 16 GB68% usedfrom $0.196/hRent

RTX A4000 16 GB68% usedfrom $0.204/hRent

RTX 4060 Ti 16 GB68% usedfrom $0.254/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 14B need?

In Q4_K_M with an 8k context, 10.7 GB: weights 8.4 GB, context cache 1.3 GB and about 1 GB of runtime memory. In BF16, 29.8 GB.

Does Qwen3 14B run on an RTX 4090 with 24 GB?

Yes. In Q4_K_M the model takes 10.7 GB and fits in 24 GB.

How much does it cost to run Qwen3 14B?

The cheapest fitting server right now is RTX 3060 12 GB, from $0.084/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.