Hardware requirements

Match Gemma 4 to the hardware you actually have.

The most useful hardware question is not “what is the biggest model?” It is “what is the biggest model that still feels sane on this device?”

Open model picker Memory troubleshooting
Hardware class
Start model
Best path
Watch for

Android phones and tablets

Android guide
E2B first, then E4B
AI Edge Gallery
CPU fallback, weak acceleration, and slower large-model tests.

iPhone and iPad

iPhone and iPad guide
E2B first
AI Edge Gallery
App-level support gaps and heavier models that load poorly.

16GB Apple silicon Mac

16GB Mac guide
E4B first, 26B A4B only if setup is disciplined
LM Studio first, llama.cpp later
Swap pressure, long-load pain, and context settings that erase headroom.

RTX desktop with 16GB VRAM

Memory guide
26B A4B as a tuned target
LM Studio baseline or llama.cpp
KV cache growth, vision overhead, and aggressive context length.

High-end desktop or workstation

Model picker
31B when convenience matters less than quality
LM Studio, llama.cpp, or an API-first stack
Over-downloading before you confirm the runtime and workload.

What moves the fit

Hardware fit is not just RAM or VRAM on paper.

Runtime overhead, context length, KV cache growth, and vision layers all change whether a setup stays usable.

That is why the same machine can feel fine in a short chat and bad in a long or tool-heavy session.

Out of memory

Start here when the model will not load, crashes, or becomes unusably slow.

Slow loading

Use this when the model technically fits but the experience still feels bad.

GPU or NPU not working

Use this when the device falls back to CPU or acceleration looks missing.

Next choices

Once the hardware class is clear, the next click should be specific.

Download

Official installers, app-store links, and the fastest safe path by device.

Try Online

Test Gemma 4 in the browser before you commit to a local setup.

Compare

Compare model sizes and runtimes without jumping between disconnected pages.