OPEN MODELS / Vision language models
Run gemma-3-4b-it-qat-q4_0-gguf on a cloud GPU
gemma-3-4b-it-qat-q4_0-gguf is a Vision language models model by google. See how much GPU memory it needs, which GPUs it fits on today and what it costs to run it yourself, billed per second.
GPUs it fits on today
Our estimate is about 4 GB for inference with GGUF weights, including working memory. Real use depends on resolution, batch size and context length.
| GPU | Memory | Per hour | 10 hours | |
|---|---|---|---|---|
| GTX 1660 S | 6 GB | $0.13 | $1.30 | Rent |
| GTX 1070 | 8 GB | $0.16 | $1.60 | Rent |
| GTX 1660 | 6 GB | $0.16 | $1.60 | Rent |
| GTX 1070 Ti | 8 GB | $0.18 | $1.80 | Rent |
| GTX 1080 | 8 GB | $0.18 | $1.80 | Rent |
Live starting prices from our catalog. You pay per second while the machine runs.
How to run gemma-3-4b-it-qat-q4_0-gguf
- 01
Rent a GPU
Pick one of the GPUs above. Billing is per second; stop the machine and you stop paying.
- 02
Open a terminal or Jupyter
Connect over SSH or open Jupyter in your browser. PyTorch and CUDA are ready to use.
- 03
Download and run the model
Run the commands below. The weights download straight from Hugging Face to your machine.
This model is gated: accept its license on Hugging Face and run huggingface-cli login before downloading.
pip install -U "huggingface_hub[cli]"
huggingface-cli download google/gemma-3-4b-it-qat-q4_0-gguf --local-dir ./model --include "*Q4_K_M*"
# llama.cpp: ./llama-server -m ./model/*.gguf --host 0.0.0.0 --port 8000Why run it yourself?
Run document and image understanding on your own GPU: no per-image fee, and files stay on your machine.
Questions
- How much GPU memory does gemma-3-4b-it-qat-q4_0-gguf need?
- By our estimate about 4 GB for inference (3.9B parameters). Higher resolution, longer context or bigger batches need more.
- How much does it cost to run gemma-3-4b-it-qat-q4_0-gguf?
- From $0.13 per hour on GTX 1660 S, billed per second. Ten hours cost about $1.30.
- Do I need a subscription?
- No. You pay only for the time your machine runs, from your USD balance. Stop it whenever you are done.
Try it on your own GPU in minutes
Pick a GPU, open Jupyter or SSH and run any open model. You pay per second and can stop anytime; no subscription, no setup fee.
- Per-second billing in USD
- PyTorch and CUDA ready in minutes
- Rent from ChatGPT or Claude too