Local AI hardware guide

What machines can run Google Gemma locally?

Updated July 28, 2026 · HiRAM Deals

A 32 GB computer is the practical starting point for experimenting with current quantized Gemma models and large contexts. Move to 64 GB or more when you want larger models, more context headroom, or several workloads at once.

Our recommendation: 32 GB RAM for smaller quantized models, 64 GB for a comfortable workstation, and 96–128 GB for larger models or lighter quantization. A supported discrete GPU improves speed. CPU-only inference works, but runs more slowly.

Which Gemma should you run?

Google’s current Gemma 4 family includes smaller edge models plus 12B, 26B mixture-of-experts, and 31B workstation models. Google lists 128K context for the smaller models and 256K for medium models. Ollama packages current quantized variants from roughly 7 GB to 20 GB before context and application overhead.

32 GB RAM

Good entry point for Gemma 4 E2B, E4B, and 12B quantized builds. It leaves useful room beyond model weights, but very long prompts can still exhaust memory.

64 GB RAM

Our general recommendation for a smooth local workstation: larger context caches, the 26B or 31B quantized models, and room for development tools.

96–128 GB RAM

Best for long contexts, multiple models, CPU-heavy inference, or larger quantizations. More RAM adds capacity. GPU bandwidth controls speed.

About the 100K-token target

A smaller quantized model can attempt a 100K-token context in 32 GB. Model weights take only part of the memory. The KV cache grows with the model and token count, while Ollama, Windows, image inputs, and open applications take their share. Choose 64 GB for regular 100K-token work, then verify the exact model, quantization, and runtime.

Install with Ollama

Ollama is the simplest cross-platform starting point. After installing it, run:

ollama run gemma4

Choose an explicit tag when you need predictable model size. See the Gemma 4 model library on Ollama.

What matters when buying a machine

Installed RAM: 32 GB minimum; 64 GB recommended for sustained local-AI work.
GPU memory: enough VRAM can keep more of the model on the GPU and materially improve speed.
Memory bandwidth: CPU-only generation often becomes bandwidth-bound.
Upgrade capacity: used workstations with spare DIMM slots can be better values than sealed consumer systems.
Storage: leave space for several multi-gigabyte model files.