Gemma 4 System Requirements
Memory is what matters. If a model fits in your computer's memory, it will run: the smallest Gemma 4 needs under 3 GB and the largest about 17.5 GB in the compact 4-bit form most people use. This page gives the figures for every size and device, a checker for your own machine, and what EmbeddingGemma 2 needs.
The four things to check
For the 4-bit builds, plus headroom for the conversation.
A graphics card or Apple silicon is much faster. The small models work on a processor alone.
For each model you download, in its 4-bit form.
Windows, macOS and Linux; Android and iOS for the phone builds.
On a PC with a graphics card, the memory that counts is the card's own. On an Apple silicon Mac, it is the computer's total memory, which the processor and graphics chip share.
Can my machine run it?
Choose your device, drag the slider to your memory, and see which models fit.
This site's rule of thumb, built on Google's memory figures. "Runs comfortably" means the Ollama download is under about 70% of your memory. "Tight fit" means the model loads with about 1.5 GB to spare. Phone results are a rough guide based on the mobile-format memory figures. Nothing here is a guarantee; the VRAM guide goes into more detail.
What each size needs
Memory for the model itself at three levels of compression. 4-bit is the usual choice.
| Model | 4-bit | 8-bit | Full precision | Ollama download | Comfortable hardware for 4-bit |
|---|---|---|---|---|---|
| Gemma 4 E2B | 2.9 GB | 5.7 GB | 11.4 GB | 4.3 GB | Any 8 GB computer; most recent phones |
| Gemma 4 E4B | 4.5 GB | 8.9 GB | 17.9 GB | 6.1 GB | An 8 GB laptop, or a 6 to 8 GB graphics card |
| Gemma 4 12B | 6.7 GB | 13.4 GB | 26.7 GB | 7.2 GB | A 16 GB laptop or Mac, or a 12 GB graphics card |
| Gemma 4 26B A4B | 14.4 GB | 28.8 GB | 57.7 GB | 16 GB | A 24 GB graphics card, or a 32 GB Mac |
| Gemma 4 31B | 17.5 GB | 34.9 GB | 69.9 GB | 19 GB | A 24 GB graphics card, or a 32 GB+ Mac |
Conversations cost memory
The figures cover the model only. Long prompts and chats need more on top, so a full 256K context takes far more than a short exchange.
The 26B A4B loads in full
Only about 3.8B parameters are active per token, but all 25.2B must sit in memory. It is quick to run, not small to load.
Images and audio add a little
Every size reads images; E2B, E4B and 12B also take audio. Each image or clip uses part of the context.
Memory figures from Google's Gemma 4 model card; Ollama sizes from its tag list. "Comfortable hardware" is this site's guidance: the model plus room for the operating system and a working conversation. Full detail is on the specs page.
Requirements for your kind of device
Windows or Linux with a graphics card
The model should fit in the card's memory: 8 GB for E4B, 12 GB for the 12B, 24 GB for the 26B A4B or 31B. If it does not fit, the rest spills on to the processor and replies slow down sharply.
Mac with Apple silicon
Memory is shared, so the model can use most of it: 16 GB suits the 12B, 32 GB the 26B A4B, and 32 GB or more the 31B. Google's AI Edge Gallery lists an M1 chip or later with macOS 14.
Guide: Mac setup.
Computer without a graphics card
E2B and E4B run acceptably on a modern processor with 8 GB of memory. Larger sizes will load if memory allows, but expect slow replies. LM Studio lists 4 GB of memory as its minimum for the smallest Gemma 4.
Guide: run locally.
Phone or tablet
E2B and E4B only, through Google AI Edge Gallery: Android 12 or later or iOS 17 or later. Downloads are 2.58 GB and 3.66 GB; the models use about 1.1 GB and 2.5 GB of memory in the mobile format.
Guides: Android, mobile downloads.
Operating-system versions from the app store listings; memory figures from Google's model card. The card-size pairings are this site's guidance. Not sure which app to use? Try the runtime selector.
Disk space
How much room each download takes. Sizes differ a little between apps because each packages the model its own way.
| Model | Ollama, QAT | Google's QAT file | LM Studio | Phone build |
|---|---|---|---|---|
| Gemma 4 E2B | 4.3 GB | 3.35 GB | 4.20 GB | 2.58 GB |
| Gemma 4 E4B | 6.1 GB | 5.15 GB | 5.90 GB | 3.66 GB |
| Gemma 4 12B | 7.2 GB | 6.98 GB | 7.40 GB | - |
| Gemma 4 26B A4B | 16 GB | 14.4 GB | 15.60 GB | - |
| Gemma 4 31B | 19 GB | 17.7 GB | 19.00 GB | - |
From Ollama's tag list, Google's QAT collection on Hugging Face, LM Studio's catalogue and Google's LiteRT-LM pages. Each app keeps its own copy, so installing two apps doubles the space. Where to get the files is on the download hub and in how to download.
What you need installed
| To do this | You need | Runs on |
|---|---|---|
| Chat on a computer | Ollama or LM Studio. Nothing else. | Windows, macOS, Linux |
| Chat on a phone | Google AI Edge Gallery | Android 12+, iOS 17+ |
| Write Python against it | Python with the Transformers library, or the ollama or openai package pointed at a local app | Any system with Python |
| Work on a Mac in Python | pip install mlx mlx-lm mlx-vlm | Apple silicon |
| Serve many users | vLLM 0.25.0 or newer | Linux servers |
| Search with EmbeddingGemma 2 | Ollama, or sentence-transformers 6.1.0 or newer | Windows, macOS, Linux |
vLLM version from the vLLM Gemma 4 recipes; sentence-transformers version from the EmbeddingGemma 2 model card. No account or licence key is needed: the models are Apache 2.0. Usage is covered in how to use and on the API page.
Server requirements
Serving other people usually means running at full precision, which needs far more memory than a 4-bit build on a laptop.
One graphics card with at least 24 GB.
One 80 GB card, such as an A100 or H100.
One 80 GB card, or two for more room.
From the hardware list in the vLLM Gemma 4 recipes. Google also publishes vLLM-ready 4-bit builds for E2B, E4B, 12B and 31B, which need much less. Setup, costs and security are on the self-hosting page; hosted alternatives under pricing.
EmbeddingGemma 2 system requirements
EmbeddingGemma 2 is Google's small model for search across text, code, images, audio and video. Its requirements are modest: it is a fraction of the size of Gemma 4, runs on a phone, and does not need a graphics card.
In use on a Pixel 11 Pro, as reported by Google.
Same device, text-only mode.
From the smallest phone build to the full Ollama download.
Useful for indexing large collections quickly.
| Build | Handles | Download | Runs on |
|---|---|---|---|
Ollama, embeddinggemma-2:270m | Text and code | 378 MB | Windows, macOS, Linux |
Ollama, embeddinggemma-2 | Text and images | 1.3 GB | Windows, macOS, Linux |
| Python, full weights | Text, code, images, audio, video | 740M parameters | Any system with Python |
| Phone builds | Text; text and vision; everything | 165, 388, 485 MB | Android, iOS |
Budget for storage, not memory
The vectors you create need space: about 1.5 GB for every million at full length, or 250 MB at the shortest.
Software
sentence-transformers 6.1.0 or newer for the full feature set. Use bfloat16 or float32, not float16.
For audio and video, use Python
Ollama lists text and image input only. The Python library covers every input type.
From Google's EmbeddingGemma 2 model card, developer guide and on-device post; Ollama sizes from its library. Storage costs are worked out on the pricing page.
What are the minimum requirements for Gemma 4?
About 2.9 GB of free memory for the smallest size, E2B, at 4-bit, and a few gigabytes of disk space. An 8 GB computer or a recent phone is enough.
How much RAM do I need?
8 GB for E2B or E4B, 16 GB for the 12B, and 32 GB for the 26B A4B or 31B on a Mac. On a PC with a graphics card, it is the card's memory that counts.
Do I need a graphics card?
No, but replies are much faster with one. Without it, stick to E2B or E4B.
How much VRAM does the 31B need?
About 17.5 GB at 4-bit, so a 24 GB card. At full precision it needs 69.9 GB.
Can I run Gemma 4 on a laptop?
Yes. A 16 GB laptop runs the 12B; an 8 GB one runs E4B. See the 12B page and all variants.
Does it work without internet?
Yes, after the first download.
What does EmbeddingGemma 2 need?
Very little: under 1.5 GB of disk space and a few hundred megabytes of memory. It runs on phones and ordinary laptops.
What if my machine is too small?
Use a hosted version: chat online, try the playground with a smaller model, or see the API page.
Your machine is ready?
Download a model and start in one command.