google

Gemma 4 31B

Google DeepMind's 31B dense multimodal model (text and image input) with a 256K-token context window.image text to textapache-2.07.1M downloadsHugging Face

Cheapest GPU in stock that fits

A single GPU with at least 75 GB for inference.

NosanaIn stock · 21 freeStart right away

1× RTX PRO 600096 GB VRAM

Nosana network, hosts worldwide

Details

Stock and price checked 16 min ago.

Self-service: you start it yourself, with no order or sales contact.

Confidential computing: data stays encrypted while the GPU works on it, so the host can't read it.

nvidia-pro-6000-cc

Full price

€2.09

per hour, whole instance, 1 GPU

View offer

GPU memory it needs

From model cards and vendor documentation. Lower precision needs less memory, with slightly lower quality. Estimates are marked.

What you want to do	Precision	Minimum	Setup	Basis
Run the model vLLM recipe minimum for this repo's BF16 checkpoint.	BF16, full	75 GB	1x 80 GB GPU (per vLLM recipe)	Published
Run the model vLLM recipe minimum for separate repo RedHatAI/gemma-4-31B-it-FP8-dynamic.	FP8, 8-bit	38 GB	—	Published
Run the model vLLM recipe minimum for separate repo nvidia/gemma-4-31B-it-NVFP4.	4-bit	19 GB	Blackwell GPU, per vLLM recipe	Published
LoRA fine-tuning Arithmetic from BF16 weight size (58.25 GiB = 31.27B parameters x 2 bytes): weights x 1.3, our standard estimate. Excludes activations beyond the overhead factor; batch size not assumed.	BF16, full	75.7 GB	—	Estimated
Full fine-tuning Arithmetic from BF16 weight size (58.25 GiB = 31.27B parameters x 2 bytes): weights x 4 states x 1.2, our standard estimate. Excludes activations beyond the overhead factor; batch size not assumed.	BF16, full	279.6 GB	—	Estimated