Install Ollama and run ollama run gemma4, which pulls Gemma 4 E4B, a 6.6GB to 9.5GB download that reads text and images with a 128K context window. Pick the size by your Mac's memory: Gemma 4 E2B (4.6GB to 7.5GB) is tight on 8 GB, where Gemma 3 4b (3.3GB) fits better; the E4B suits 16 GB, where the 12b (7.7GB to 8.0GB) is Tight and a good fit from 24 GB; the 26b, a mixture of experts with 4B active parameters (16GB to 19GB), is Tight on 32 GB and a good fit from 48 GB; and the 31b (19GB to 20GB) wants 48 GB or more. The E in E2B and E4B means effective parameters, sized for laptops and edge devices. Gemma 4 is Apache 2.0 licensed; Gemma 3 ships under Google's Gemma terms. Once it runs, point any app that speaks Ollama at it, or install Grux AI (MIT), whose model cookbook lists Gemma 4 12b and 26b by memory and can use them to act on your mail, calendar, files and shell.
brew install ollama
ollama run gemma4 # Gemma 4 E4B, 6.6GB to 9.5GB
ollama run gemma4:12b # 7.7GB to 8.0GB, for 24 GB Macs, tight on 16 GB
ollama run gemma4:26b # 16GB to 19GB, for 48 GB or more
| Your Mac's memory | Gemma model | Download |
|---|---|---|
| 8 GB | Gemma 3 4b (Gemma 4 E2B is tight) | 3.3GB |
| 16 GB | Gemma 4 E4B; the 12b is Tight | 6.6GB to 9.5GB |
| 24 GB | Gemma 4 12b, a good fit | 7.7GB to 8.0GB |
| 32 GB | Gemma 4 26b is Tight; 12b is a great fit | 16GB to 19GB |
| 48 GB | Gemma 4 26b, a good fit, or 31b | 16GB to 20GB |
| 64 GB or more | Gemma 4 31b comfortably | 19GB to 20GB |
Where a row rates a model in Grux AI's catalogue (Gemma 4 12b and 26b), it uses the rating from the which LLM can my Mac run page, which rates by the memory a model uses while running, against about three quarters of your Mac's memory. Other rows are a rough guide from file sizes. Sizes are from Ollama's Gemma 4 and Gemma 3 pages, read on 7 October 2026.
Gemma 4 is the newer family and the one to start with: E2B and E4B for laptops, a 12b, a 26b mixture of experts and a 31b dense model, all reading images as well as text. Ollama's Gemma 4 page lists the 31b ahead of Gemma 3 27b on MMLU Pro, AIME 2026 and LiveCodeBench. Gemma 3 still has the smallest sizes, down to 270m and 1b, which suit an 8 GB Mac or a quick test.
| Claim | Source |
|---|---|
| Gemma 4 sizes E2B to 31b, 128K and 256K context, text and image, the E means effective | Ollama library, gemma4 |
| Gemma 3 sizes from 270m to 27b | Ollama library, gemma3 |
| Gemma 4 is Apache 2.0 | Gemma 4 12B on Hugging Face |
| Gemma 3 is under Google's Gemma terms | Gemma 3 4B on Hugging Face |
| Grux's cookbook lists Gemma 4 12b and 26b by memory | Cookbook.swift |
ollama run gemma4 pulls by default, which leaves room for context on 16 GB. The Gemma 4 12b runs there but is Tight; it is a good fit from 24 GB.