What GPU do I need to run google/gemma-4-31B?
32.7B parameters, published in BF16. View on Hugging Face
Set up gemma-4-31Bgemma-4-31B is published by google on Hugging Face, with 717,273 downloads and 516 likes to date. It's a Gemma4ForConditionalGeneration model built for image-text-to-text, published natively in BF16.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| BF16 | 60.9 GB | 73.1 GB | A100 | 1 | $1.16/hr |
| cheaper alt. | RTX A5000 | 4 | $0.704/hr | ||
| FP8 (quantized) | 30.4 GB | 36.5 GB | L40 | 1 | $0.759/hr |
| cheaper alt. | RTX 4000 SFF Ada | 2 | $0.396/hr | ||
| INT4 (quantized) | 15.2 GB | 18.3 GB | RTX A5000 | 1 | $0.176/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run gemma-4-31B at its published (BF16) precision: 1× A100, at $1.16/hr per GPU ($1.16/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
gemma-4-31B: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Sliding-window attention: windowed layers stop growing at the window, the full-attention layers keep growing. Cached per token: 983,040 bytes at 16-bit while the context is inside the window.
| Context | KV cache, one sequence |
|---|---|
| 4K | 1.4 GB |
| 32K | 5.8 GB |
| 128K | 20.8 GB |
| 256K (model maximum) | 40.8 GB |
Computed from the layer, head and window counts in the model's own configuration (google/gemma-4-31B), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
gemma-4-31B: common questions
Can gemma-4-31B run on a single GPU?
Yes, but not on a desktop card. At BF16 it needs 73.1 GB of VRAM, more than a single 32 GB desktop card holds. The cheapest live single-card fit is a 80.0 GB A100 at $1.16/hr.
What is the least VRAM gemma-4-31B can run in?
18.3 GB, at INT4 (quantized), which fits a 24 GB card, against 73.1 GB at BF16. That row assumes an INT4 checkpoint has actually been published for this model, so check its Hugging Face page before planning around it.
Does quantizing gemma-4-31B lower the GPU bill?
Yes. At BF16 the cheapest live fit is one A100 at $1.16/hr. At INT4 (quantized) it drops to one RTX A5000 at $0.176/hr, provided a quantized checkpoint exists for it.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Gemma 4 models
- gemma-4-31B-it-uncensored (32.7B, BF16)
- gemma-4-31B-it (31.3B, BF16)
- gemma-4-31B-it-FP8-block (31.3B, F8_E4M3)
- gemma-4-31B-it-FP8-dynamic (31.3B, F8_E4M3)
- Artemis-31B-v1.2 (31.3B, BF16)
Lineage
Quantized and fine-tuned versions
Alternatives at this size
Other models for image-text-to-text within about a third of gemma-4-31B's 32.7B parameters, from other model lines.
- Qwen3.6-27B (27.8B, BF16)
- diffusiongemma-26B-A4B-it (25.8B, BF16)
- Muse-Glimmer-30B (29.8B, BF16)
- Infinity-Parser2-Pro (35.1B, BF16)
- InternVL2-26B (25.5B, BF16)
More on gemma-4-31B
Fits on an 80 GB GPU at BF16: every model that fits in 80 GB.
Fits on a 48 GB GPU at FP8: every model that fits in 48 GB.
Fits on a 24 GB GPU at INT4: every model that fits in 24 GB.
Best vision-language models: how gemma-4-31B ranks against the rest.
Related reading: A100 pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
