Install Ollama and run ollama run llama3.1, which pulls the 8b, a 4.9GB download with a 128K context window. Pick the size by your Mac's memory: Llama 3.2 3b (2.0GB) suits an 8 GB Mac, where it is a good fit; the 3.1 8b is the usual choice from 16 GB up but is not in the app's catalogue, so it is not rated; and the 70b, Llama 3.3 (43GB), is rated by the memory it needs while running, 60 GB in the app's catalogue, against three quarters of your Mac's memory: too big on 64 GB, tight on 96 GB and a good fit on 128 GB. Llama 4 Scout (67GB, 16 experts) is not in that catalogue, so it is not rated here; Maverick (245GB) is out of reach of nearly every Mac. Llama is not open source in the usual sense: each release ships under Meta's own Llama Community License. Once it runs, any app that speaks Ollama can use it, including Grux AI (MIT), which can use it to act on your mail, calendar, files and shell.
brew install ollama
ollama run llama3.2:3b # 2.0GB, for 8 GB Macs
ollama run llama3.1 # the 8b, 4.9GB, for 16 GB and up
ollama run llama3.3 # the 70b, 43GB, for 96 GB or more
| Your Mac's memory | Llama model | Download |
|---|---|---|
| 8 GB | Llama 3.2 3b, a good fit (Llama 3.1 8b is not rated) | 2.0GB |
| 16 GB | Llama 3.1 8b (not rated) | 4.9GB |
| 24 GB or 32 GB | Llama 3.1 8b (not rated); the 70b is too big | 4.9GB |
| 48 GB | Llama 3.1 8b (not rated); the 70b is too big | 4.9GB |
| 64 GB | The 70b is too big | 43GB |
| 96 GB | Llama 3.3 70b, tight | 43GB |
| 128 GB | Llama 3.3 70b, a good fit; Llama 4 Scout is not rated | 43GB, 67GB |
The table uses one rule, the one on the which LLM can my Mac run page: a model is rated by the memory it needs while it is serving, against about three quarters of unified memory. Only two Llama models are in the app's catalogue, read at commit 4b8d1b6: Llama 3.2 3b (3 GB while serving) and Llama 3.3 70b (60 GB). Those are the rated rows. Llama 3.1 8b and 70b and Llama 4 Scout are not in it, so this page gives no rating for them: the 3.1 70b is the same 43GB download as the 3.3 but is not itself rated, and for Scout the only figure is its 67GB download. Download sizes are from Ollama's Llama 3.2, Llama 3.1, Llama 3.3 and Llama 4 pages, read on 18 November 2026.
Llama is not published under an open source licence. Each release has its own Meta licence, for example the Llama 3.1 Community License Agreement, so read the terms before you build something on it. For a model under a standard open source licence, see Qwen or gpt-oss.
| Claim | Source |
|---|---|
| Llama 3.2 sizes 1b and 3b, and their downloads | Ollama library, llama3.2 |
| Llama 3.1 sizes 8b, 70b and 405b, and their downloads | Ollama library, llama3.1 |
| Llama 3.3 is a 70b model, 43GB | Ollama library, llama3.3 |
| Llama 4 Scout and Maverick, multimodal mixture of experts, with their downloads | Ollama library, llama4 |
| Llama 3.1 ships under Meta's own community licence | Meta, Llama 3.1 Community License Agreement |
ollama run llama3.1 pulls by default. It is a 4.9GB download with a 128K context window and leaves room for context on 16 GB.