Llama 3.2 1B and 3B on iPhone and Mac: what they are and how to run them
· By Modelbase Team
Llama 3.2 1B and 3B are Meta’s smallest Llama models: text-only models built to run on phones and other edge devices. Meta released them in September 2024, made them by shrinking and distilling the larger Llama 3.1 models, and aimed them at on-device jobs like summarizing, following instructions and rewriting (Meta). In 4-bit form, Llama 3.2 (1B) is a 713 MB download and Llama 3.2 (3B) is 1.82 GB, small enough to run offline on an iPhone with an app like Modelbase.
Llama 3.2 at a glance
| Llama 3.2 1B | Llama 3.2 3B | |
|---|---|---|
| Parameters | 1.23 billion | 3.2 billion |
| Input | Text only | Text only |
| Context length | 128K tokens | 128K tokens |
| Official languages | English, German, French, Italian, Portuguese, Hindi, Spanish, Thai | Same |
| Knowledge cutoff | December 2023 | December 2023 |
| Released | 25 September 2024 | 25 September 2024 |
| License | Llama 3.2 Community License | Same |
| Download in Modelbase (4-bit) | 713 MB | 1.82 GB |
Sources: Meta, Llama 3.2 1B model card; download sizes from Modelbase’s catalog.
How Meta made Llama 3.2 small
Meta didn’t train the 1B and 3B models from scratch. It started from Llama 3.1 8B, removed parts of the network in one pass (“structured pruning”), and then trained the smaller models using the outputs of Llama 3.1 8B and 70B as targets, a technique called knowledge distillation (Meta). The idea is that a small model learns more from a big model’s answers than from raw text alone.
What about Llama 3.2 Vision?
Llama 3.2 also has two vision models, 11B and 90B, which can read images. They’re far too large for a phone, and the 1B and 3B models can’t see images at all. On an iPhone, use a small vision model instead: Qwen 3.5, Gemma 4 (E2B) or LFM 2.5 VL in Modelbase all answer questions about photos on the device. See how to ask AI about a photo on your iPhone.
What Llama 3.2 1B and 3B are good at
Meta’s model card lists uses like mobile writing assistants and query and prompt rewriting (Meta), and Modelbase’s catalog describes the models as good for multilingual dialogue and summarization. In everyday use that means:
- Rewriting: make an email more polite, shorten a paragraph, fix the tone of a message.
- Summarizing: the gist of an article you paste in, or of a long message.
- Conversation and brainstorming: names, gift ideas, plans, first drafts.
- Translation between its supported languages.
And where to be careful:
- Facts after December 2023 aren’t in the model. Turn on web search for anything recent.
- Obscure facts in general: a 1B or 3B model holds far less knowledge than a cloud chatbot. Check names, numbers and dates.
- Math and logic: a reasoning model does better; see what a reasoning model is.
1B or 3B?
- Llama 3.2 (1B) is quick and light. Modelbase recommends it for iPhone 15 and newer. Use it for short rewrites, summaries and quick questions.
- Llama 3.2 (3B) has more than twice the parameters, so expect more complete answers, at the cost of memory. Modelbase recommends it for iPhone 15 Pro and newer, and it runs comfortably on any Apple silicon Mac.
If your iPhone is older, the app’s model cards say which models suit it; LFM 2.5 (350M) and Qwen 3 (0.6B) are the light options.
How to run Llama 3.2 on iPhone, iPad and Mac
Modelbase is a free app that runs open models on iPhone and iPad (iOS 18 or later) and Macs with Apple silicon (macOS 15 or later), with Apple’s MLX. No account, no Terminal.
- Install Modelbase from the App Store.
- Open the catalog. Tap the model’s name at the top of the chat, then Manage models > Local.
- Search for “Llama” and tap Download on Llama 3.2 (1B) or Llama 3.2 (3B). Download on Wi-Fi if you can.
- Pick it in the model picker and start chatting. It works offline from then on.
Llama 3.2’s 128K context on a phone
Llama 3.2 supports 128K tokens of context, but a phone can’t hold that much conversation in memory. Modelbase sets the context window by your device’s memory (2,048 to 8,192 tokens on iPhones, 8,192 on a Mac by default) and summarizes earlier messages when a chat grows longer. Our guide to tokens and context windows explains why.
Dolphin Llama 3.2
The catalog also has Dolphin Llama 3.2 (1B and 3B), community fine-tunes of Llama 3.2 by Dolphin AI. They’re the same size as Meta’s models; the catalog describes them as general-purpose models for coding, math and everyday use.
Prompts that suit Llama 3.2
Small models do best with short, specific requests. Some that work well:
- “Rewrite this message so it sounds friendly but firm: …”
- “Summarize this article in three bullet points: …”
- “Give me five name ideas for a bakery that sells sourdough and coffee.”
- “Translate into Spanish, keeping it informal: …”
- “Turn these notes into a short email to my team: …”
A few habits help: paste the text you want worked on instead of describing it, say how long the answer should be, and start a new chat when you change topic, so the model’s small context window holds only what matters. If an answer looks wrong, ask the same question with web search on.
Llama 3.2 offline, on the go
Because the model runs on the device, everything above works in airplane mode or with no signal: download Llama 3.2 on Wi-Fi before you travel and it’s there when you need it. Our guide to using AI offline lists what else works without a connection.
Llama 3.2 vs newer small models
Llama 3.2 came out in September 2024, and newer small model families have arrived since: Qwen 3.5, Gemma 4, LFM 2.5. They add things Llama 3.2 1B and 3B don’t have, such as image input and an optional thinking mode (Qwen 3.5). Here’s how the sizes compare in Modelbase:
| Model | Images | Thinking | Download |
|---|---|---|---|
| Llama 3.2 (1B) | No | No | 713 MB |
| Llama 3.2 (3B) | No | No | 1.82 GB |
| LFM 2.5 (1.2B), included | No | No | 663 MB |
| Qwen 3.5 (2B) | Yes | Optional | 1.75 GB |
| Phi 4 Mini (3.8B) | No | No | 2.18 GB |
Llama 3.2 remains a sensible choice for plain-text writing and conversation in its eight languages. For a closer comparison, read Gemma vs Llama vs Qwen, and for the general picture, which model to run on your iPhone and how to run an LLM locally.
The Llama 3.2 license in one paragraph
Llama 3.2 is released under the Llama 3.2 Community License, Meta’s own license, not a standard open-source one (Meta). It allows commercial use under its terms and an acceptable use policy. One clause often comes up: for the multimodal (vision) models, the license’s rights aren’t granted to individuals living in the EU. That clause is about the vision models; the 1B and 3B are text-only.
FAQ
What is Llama 3.2 1B?
Meta’s smallest Llama model: a text-only model with about 1.23 billion parameters and a 128K-token context, released in September 2024 for on-device use. In Modelbase it’s a 713 MB download.
Can Llama 3.2 run on an iPhone?
Yes. Llama 3.2 (1B) runs on iPhone 15 and newer and Llama 3.2 (3B) on iPhone 15 Pro and newer, offline, in Modelbase.
What is the context window of Llama 3.2?
128K tokens for the 1B and 3B models. On a phone, apps use a smaller window to fit in memory.
Can Llama 3.2 3B see images?
No. Llama 3.2 1B and 3B are text-only. The vision models are 11B and 90B, too large for a phone.
What languages does Llama 3.2 support?
Officially English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
Is Llama 3.2 free?
Yes, the models are free to download under Meta’s Llama 3.2 Community License, and Modelbase is free.
Try it on your iPhone, iPad or Mac
Modelbase is free on the App Store.