Gemma 4 / official memory worksheet

6.7 GB for 12B Q4. 69.9 GB for 31B BF16.

Q: How much GPU memory does Gemma 4 12B need?

Google lists approximately 26.7 GB at BF16, 13.4 GB at SFP8, and 6.7 GB at Q4_0. The official BF16 Hugging Face checkpoint is 23,919,549,408 bytes, or 22.28 GiB, before runtime memory and context state.

Q: Can Gemma 4 12B run on an 8 GB GPU?

The official Q4_0 planning figure is 6.7 GB and the official QAT Q4_0 GGUF files total 6.66 GiB. An 8 GB card leaves very little room for runtime workspace, context memory, and platform differences, so treat it as a tight loading experiment rather than a production recommendation.

Q: How much GPU memory does Gemma 4 26B A4B need?

Google lists approximately 57.7 GB at BF16, 28.8 GB at SFP8, and 14.4 GB at Q4_0. Although only about 4B parameters are active per token, all expert weights still need resident storage.

Q: How much GPU memory does Gemma 4 31B need?

Google lists approximately 69.9 GB at BF16, 34.9 GB at SFP8, and 17.5 GB at Q4_0. Those are inference-loading estimates; long context and concurrency need additional memory.

Q: Do Gemma 4 memory estimates include KV cache?

Google describes its table as approximate memory required to load the model with 20 percent overhead. It is not a guarantee that a specific context length, batch size, multimodal input, or concurrent workload fits. Measure cache and runtime allocations in the chosen serving stack.

Google publishes one inference-memory table for the full Gemma 4 family. Use it as the loading baseline, then add context, concurrency, multimodal inputs, and runtime-specific reserve.

Open the editable Gemma 4 12B Q4 plan →

Quick answer / official loading estimates

The 12B model spans consumer GPUs. The 31B model does not at BF16.

Google's values are approximate inference memory, not raw parameter arithmetic. The editable links preserve the published checkpoint or family estimate as the starting input so the runtime reserve stays visible.

12B · Q4_0 · 8 GB card TIGHT FIT 6.66 GiB artifact before cache and runtime variance → 12B · Q4_0 · 16 GB card ROOM TO MEASURE Artifact plus a visible generic runtime reserve → 12B · BF16 · 32 GB card 26.74 GiB + cache Exact files plus 20% reserve; context excluded → 31B · Q4_0 · 24 GB card 17.5 GB baseline Published loading estimate; validate the runtime →

Official Gemma 4 inference table

Five sizes. Three precisions. One loading baseline.

Google says the table includes approximately 20% overhead for model loading. Treat every value as a first gate, then verify the exact runtime, context, batch size, modalities, and concurrency.

Swipe the matrix horizontally →

Gemma 4 model	BF16	SFP8	Q4_0	Deployment boundary
E2B	11.4 GB	5.7 GB	2.9 GB	PLE makes total storage larger than 2.3B effective parameters imply
E4B	17.9 GB	8.9 GB	4.5 GB	8B parameters with embeddings, 4.5B effective
12B	26.7 GB	13.4 GB	6.7 GB	Dense 11.95B; official BF16 files total 22.28 GiB
26B A4B	57.7 GB	28.8 GB	14.4 GB	All 25.2B weights reside even though 3.8B activate
31B	69.9 GB	34.9 GB	17.5 GB	Dense 30.7B; long context still adds state

Source: Google AI for Developers, “Gemma 4 model overview,” retrieved August 7, 2026. Values may change with inference tool and environment.

Read the estimate

Effective is not resident.
E-model embeddings and MoE experts still consume memory even when fewer parameters drive each token.

Loading is not serving.
Context, batch size, multimodal encoders, kernels, and fragmentation can move the fit.

Eight gigabytes is a test.
The 12B Q4 artifact can be tight enough that runtime choice decides whether it loads.

12B artifact cross-check

The official files explain the published 12B floor.

BF16 checkpoint
23,919,549,408 bytes · 22.28 GiB

QAT Q4_0 GGUF
7,150,994,912 bytes · 6.66 GiB

Google loading estimates
26.7 GB BF16 · 6.7 GB Q4_0

The BF16 file total plus a 20% reserve is 26.73 GiB, closely matching Google's 26.7 loading estimate. The Q4_0 GGUF artifact itself is already 6.66 GiB, so an 8 GB device should be treated as a tight experiment, not guaranteed capacity.

Primary-source inputs

Start with Google's table and exact model files.

The family matrix and overhead note come from Google AI for Developers. Architecture and file totals come from Google's official Hugging Face repositories.

Open the GPU memory calculator → Compare DiffusionGemma hardware → Find models for your GPU budget → Official Gemma 4 model overview → Official Gemma 4 12B model card → Official 12B BF16 checkpoint → Official 12B QAT Q4_0 files →

How much GPU memory does Gemma 4 12B need?

Google lists about 26.7 GB at BF16, 13.4 GB at SFP8, and 6.7 GB at Q4_0. The official BF16 files total 22.28 GiB before runtime and context state.

Can Gemma 4 12B run on an 8 GB GPU?

The official QAT Q4_0 GGUF files total 6.66 GiB. That leaves little margin on an 8 GB card, so runtime, context length, offload behavior, and fragmentation decide whether the workload is usable.

How much GPU memory does Gemma 4 26B A4B need?

Google lists 57.7 GB at BF16, 28.8 GB at SFP8, and 14.4 GB at Q4_0. Active parameters reduce compute, not the storage required for all expert weights.

How much GPU memory does Gemma 4 31B need?

Google lists 69.9 GB at BF16, 34.9 GB at SFP8, and 17.5 GB at Q4_0 before workload-specific validation.

Do Gemma 4 memory estimates include KV cache?

The table is a model-loading estimate with approximate overhead. It does not guarantee a particular context length, modality mix, batch size, or concurrency level.