How to run Gemma 4 12B locally on a 16 GB laptop

To run Gemma 4 12B on a laptop, install Ollama and type "ollama run gemma4:12b" in a terminal. The download is about 8 GB, and Google says the model runs on laptops with 16 GB of GPU memory (VRAM) or unified memory.

You will get the exact steps, the model tags from the Ollama library, what to do if your laptop is too slow, and three ways a student can use a local model for real work. If you are new to local models, read our post on small language models first.

What is Gemma 4 12B?

Gemma 4 12B is an open model from Google with 12 billion parameters. Google released it on 3 June 2026 as a dense multimodal model, which means one set of weights handles text, images and audio. Google calls its design "encoder-free": there is no separate vision or audio encoder bolted on.

The Gemma 4 family also has smaller models (E2B and E4B) and larger ones (26B and 31B). The Gemma 4 model card says the family supports over 140 languages, which matters if you want to test Hindi, Bengali or Tamil.

Can a 16 GB laptop run Gemma 4 12B?

A 16 GB laptop can run Gemma 4 12B if the 16 GB is GPU memory or unified memory. The Google Developers Blog guide says the model is "small enough to run locally on dedicated GPU laptops with 16GB VRAM or unified memory."

Read that line carefully, because many students get it wrong:

• Apple Silicon Macs: an M-series Mac with 16 GB uses unified memory, so the CPU and GPU share all 16 GB. This matches Google's line.
• Gaming laptops: a Windows laptop with 16 GB of RAM and a 6 GB or 8 GB graphics card does not have 16 GB of VRAM. Ollama can still run the model, but part of it will run on the CPU, which is slower.
• No graphics card: the model may still start on 16 GB of RAM, but expect slow replies. Try a smaller model first.

Google does not publish a speed number for laptops, and speed depends on your chip. Test it yourself before you plan a project around it.

How to run Gemma 4 with Ollama: step by step

1. 1Download Ollama for Windows, macOS or Linux from ollama.com/download and install it.
2. 2Check you have at least 10 GB of free disk space. The default Gemma 4 12B download is about 8 GB.
3. 3Open a terminal (Terminal on Mac, PowerShell on Windows).
4. 4Type "ollama run gemma4:12b" and press Enter. Ollama downloads the model the first time, then opens a chat.
5. 5Type a question and press Enter. Type "/bye" to leave the chat.
6. 6Next time, the same command starts the model straight away, with no download.

If you only want to download the model now and chat later, use "ollama pull gemma4:12b". To see what is on your laptop, use "ollama list". To delete a model and free disk space, use "ollama rm gemma4:12b".

Which Gemma 4 model tag should I use on Ollama?

The Gemma 4 page on the Ollama library lists many tags for the 12B model. As of November 2026, the main ones are:

• gemma4:12b: the default. Start here.
• gemma4:12b-it-q4_K_M: 4-bit, 8.0 GB download.
• gemma4:12b-it-qat: 4-bit trained for quantization, 7.2 GB, the smallest 12B file.
• gemma4:12b-it-q8_0: 8-bit, 13 GB. Too big for most 16 GB laptops.
• gemma4:12b-it-bf16: full precision, 24 GB. Not for a 16 GB laptop.
• gemma4:12b-mlx: 7.7 GB, a version in Apple's MLX format for Macs.

Fewer bits means a smaller file and less memory, with a small loss in quality. For a 16 GB laptop, stay with the default or a 4-bit tag.

How to run Gemma 4 in LM Studio

LM Studio is a free desktop app that runs models with buttons instead of a terminal. Google lists it as one of the ways to try Gemma 4 12B.

1. 1Download LM Studio from lmstudio.ai and install it.
2. 2Open the search (Discover) tab and search for "Gemma 4 12B".
3. 3Pick a 4-bit version that LM Studio marks as fitting your memory, and download it.
4. 4Load the model and type in the chat box.

LM Studio shows the file size before you download, so check that it leaves room in your memory for the browser and other apps.

What if Gemma 4 12B is slow on my laptop?

If Gemma 4 12B is slow or will not load, switch to a smaller Gemma 4 model. The Ollama library lists two:

• gemma4:e4b: 6.6 GB download (4-bit is 6.6 GB, QAT is 6.1 GB).
• gemma4:e2b: 4.6 GB download (QAT is 4.3 GB). Try this on 8 GB laptops.

Note that "ollama run gemma4" with no tag gives you E4B, not 12B. Other fixes that help:

• Close apps: browsers with many tabs eat memory.
• Plug in: laptops on battery often run the chip slower.
• Short prompts: very long inputs use more memory and take longer.

What is the Gemma 4 licence?

Gemma 4 is released under the Apache 2.0 licence, as the Gemma 4 model card states. Apache 2.0 lets you use, change and share the model, also for commercial work, as long as you keep the licence and copyright notices.

3 ways students can use Gemma 4 locally

• Private reading notes: paste a paper abstract or your own notes and ask for a plain summary. Nothing leaves your laptop, which helps with unpublished work.
• Language tests: ask the same 20 questions in English and in your own language, and score the answers. That small study can become your first research project.
• Offline agents: Google pitches the 12B model for local agents. Read what AI agents are, then build a simple one that sorts your files or answers from your course PDFs.

Our advice: a local model makes mistakes on facts, so never paste its claims or citations into a paper without checking the source.

At APJ-RAISE, students test open models like Gemma together and turn the results into research. Joining is free.