Gemma infrastructure

Gemma Hardware Requirements: RAM and VRAM for Gemma 3

Gemma 3 covers a useful local-inference range from 1B through 27B parameters. This page computes weight floors from Google's published family sizes, then separates those floors from the extra memory a real runtime, context cache and multimodal workload can consume.

LLM RAM and VRAM planning layers: model weight memory determined by quantization, runtime overhead and context reserve for attention and concurrency — Memory planning components. Actual peak RAM and VRAM vary by model artifact, runtime, context and concurrency; this is not benchmark data.

Gemma hardware requirements: quick answer

Gemma 3 1B and 4B fit easily on modest local hardware at 4-bit. The 12B model starts at a 6 GB 4-bit weight floor, while 27B starts at 13.5 GB before runtime, context and vision overhead.

Computed RAM / VRAM floors by model

Weight memory is parameter count × bits per weight ÷ 8. The final column is deliberately more conservative: it runs the same 4-bit estimate through AIG's default 20% runtime reserve and 2 GB context/cache reserve. Neither value includes workload-specific concurrency, unusually long context, offload buffers, or multimodal activations.

Model	16-bit floor	8-bit floor	4-bit floor	4-bit planning target
Gemma 3 1B 1B parameters	2 GB	1 GB	0.5 GB	2.6 GB
Gemma 3 4B 4B parameters	8 GB	4 GB	2 GB	4.4 GB
Gemma 3 12B 12B parameters	24 GB	12 GB	6 GB	9.2 GB
Gemma 3 27B 27B parameters	54 GB	27 GB	13.5 GB	18.2 GB

CALCULATED. The planning target is weight floor + 20% runtime reserve + 2 GB context reserve. Validate the actual artifact and runtime before buying hardware.

Recalculate Gemma for your quantization and reserves

Starts with 12B parameters. Change precision, runtime reserve or context reserve to match your deployment.

Model presetModel parameters (billions)Weight precision / quantizationRuntime / framework reserve (%)Context / cache reserve (GB)Usable RAM / VRAM budget (GB)Enter the amount available to the model after OS, apps and other workloads. For a dedicated GPU, use its usable VRAM; for Apple silicon, use available unified memory.

Why context reserve is manual: exact attention-cache memory cannot be inferred safely from parameter count alone. For a real deployment, measure the specific model and runtime at the context length and batch size you intend to use.

Planning estimate

Weight-memory floor: 6.0 GB
Runtime reserve: 1.2 GB
Context reserve: 2.0 GB

Planning total

9.2 GB

Memory-budget screen

6.8 GB preliminary headroom

This is an arithmetic screen, not a fit guarantee. Real quantized model files, cache allocations and runtime overhead may exceed these estimates. Verify the exact artifact before buying.

Weight memory is parameters × bits ÷ 8. Treat the total as a capacity-planning starting point, not a guaranteed runtime requirement.

Compare this estimate with observed model evidence

The weight-memory floor does not include actual artifact metadata or measured runtime allocations. AIG publishes reproducible Ollama runs with host, context, precision and limitations. Only use measurements from the same model and runtime conditions as supporting evidence, not as proof of this configuration.

Inspect measured benchmark records →

Your next step

This planning estimate is within a base Mac mini's advertised memory tier, but real quantized files, cache and macOS usage may still prevent the model from fitting. Check the exact configuration.

Check this against a Mac mini configuration →

Open the full RAM and VRAM calculator → for the method, the caveats, and how it differs from the artifact-based compatibility checker.

What changes for Gemma

Gemma 3 4B, 12B and 27B are multimodal models. A text-only parameter calculation does not include image-encoder activations or workload-specific cache use, so treat the planning target as a starting reserve rather than a measured peak.

Weight fit is not the same as deployment fit

Check the exact GGUF, safetensors, AWQ, GPTQ or runtime-specific artifact you will load.
Reserve memory for KV cache, runtime allocations, the operating system and concurrent requests.
Unified-memory Macs share the same pool with the OS and applications; discrete-GPU systems still need system RAM.
For MoE models, active parameters reduce compute per token but do not automatically remove inactive expert weights from memory.

Sources

Google Gemma 3 release collection— Official 1B, 4B, 12B and 27B checkpoints plus QAT/GGUF variants

Estimate model memory →Check a specific artifact →Plan a local AI server →Compare GPU cloud providers →

Agent Infrastructure Recommendation Kit

Sizing infrastructure for Gemma?

Use the Agent Infrastructure Recommendation Kit to carry these memory assumptions into a documented local-vs-cloud infrastructure decision with explicit evidence and verification steps.

Get the Recommendation Kit — $19