Gemma 3 is one of the best open models you can run at home, and it's multimodal — it reads images too. Here's how to run it in one command, and which size fits your GPU.
One command: install Ollama, then run `ollama run gemma3` — it downloads the model and starts a chat. Gemma 3 is Google's open model family and one of the best you can run at home, with a standout feature: it's multimodal, so it can look at images, not just read text. It comes in sizes from a tiny 4B up to a capable 27B, so there's a version for almost any GPU, and it runs privately and offline for free. Here's the full guide, including which size to pick and how it stacks up against Qwen3.
Step 1: install and run
The fastest path is Ollama. Install it, then pick the size that fits your GPU:
# install Ollama from ollama.com, then run the size that fits your GPU:
ollama run gemma3:4b # ~8GB VRAM — light and fast
ollama run gemma3:12b # ~12-16GB VRAM — the sweet spot
ollama run gemma3:27b # ~24GB VRAM — most capable
# first run downloads the model, then you're chatting.
# gemma3 can also read images — pass one in a supported UI like LM Studio.
Prefer a graphical app? LM Studio runs the same Gemma 3 models and makes it easy to drop an image in for the multimodal features.
Step 2: pick the right size — and use the image feature
Match the size to your [VRAM](/how-much-vram-do-you-need-to-run-ai-models-2026): Gemma 3 4B runs on an 8GB card and is surprisingly capable for its size; Gemma 3 12B is the sweet spot on a 12-16GB card and where quality really shows; Gemma 3 27B is the most capable, fitting a 24GB card. The feature that sets Gemma 3 apart from most local models is multimodality — it can look at an image and answer questions about it, describe it, or extract text, all locally and privately. That's genuinely useful and rare among open models you can run at home. If you want a text-only all-rounder, Qwen3 is its main rival and equally excellent; if you want image understanding too, Gemma 3 is the pick. Both are among the best free local models.
Which Gemma 3 size to run
8GB
Your GPU VRAM
Gemma 3 4B
Run
Light, capable
12-16GB
Your GPU VRAM
Gemma 3 12B
Run
The sweet spot
24GB
Your GPU VRAM
Gemma 3 27B
Run
Most capable
Your GPU VRAM
Run
Notes
8GB
Gemma 3 4B
Light, capable
12-16GB
Gemma 3 12B
The sweet spot
24GB
Gemma 3 27B
Most capable
Gemma 3 runs in one command and, uniquely for a local model, can also understand images. · Unsplash
Quick answers
How do I run Gemma 3 locally?
Install Ollama (free, from ollama.com), then run 'ollama run gemma3' — pick a size like gemma3:4b, gemma3:12b, or gemma3:27b to match your GPU. The first run downloads the model and starts a chat; it runs privately and offline on your own hardware, free. Gemma 3 is multimodal, so it can also analyze images — use a graphical app like LM Studio to drop images in. Match the model size to your VRAM: 4B for 8GB, 12B for 12-16GB, 27B for 24GB.
Which Gemma 3 size should I run?
Match it to your GPU's VRAM. Run Gemma 3 4B on an 8GB card (light and capable), Gemma 3 12B on a 12-16GB card (the sweet spot where quality shows), and Gemma 3 27B on a 24GB card (the most capable). Don't run a size larger than your VRAM allows, or it will spill to system RAM and slow down. If unsure, start one size below your card's maximum, confirm the speed, then try larger.
Is Gemma 3 better than Qwen3 for local AI?
They're both excellent top-tier open models, and the choice depends on your needs. Gemma 3's standout feature is multimodality — it can understand images, which most local models can't — making it the pick if you want image analysis. Qwen3 is an outstanding text all-rounder, often slightly ahead on pure text tasks and coding at comparable sizes. For text-only use, either is a great choice; for image understanding on your own hardware, Gemma 3 wins. Many people keep both installed via Ollama.
Gemma 3 is a one-command local model with a rare multimodal edge — pick the size for your VRAM and go. New to local AI? Start with the beginner's guide; compare with Qwen3 and the best free models. Source: Google and Ollama.
