Llama 3.2

What Llama 3.2 is and who makes it

Llama 3.2 is a set of models Meta released on September 25, 2024. It went in two directions at once. On one side were two very small text models, Llama 3.2 1B and 3B, designed to run on phones, laptops and other devices with limited computing power. On the other side were Meta's first Llama vision models, Llama 3.2 11B Vision and 90B Vision, which accept an image together with text and answer in text. All of them are downloadable under the Llama 3.2 Community License.

The Llama 3.2 family

Llama 3.2 1B and 3B (text)

According to Meta's model card, the 1B model has about 1.23 billion parameters, both small models support a 128K context window, and they were trained on up to 9 trillion tokens with data up to December 2023. Meta built them partly by learning from the outputs of the larger Llama 3.1 8B and 70B models, a technique called knowledge distillation. Meta lists intended uses such as assistant-like chat, knowledge retrieval, summarization, writing assistants on mobile devices, and rewriting queries and prompts.

Meta also published quantized versions of 1B and 3B, compressed to run on phone processors with less memory. Those quantized versions have a shorter context window of 8K tokens.

Llama 3.2 11B and 90B Vision

The vision models are built on top of the Llama 3.1 text model with a separately trained vision adapter, which feeds what an image encoder sees into the language model. Meta lists visual recognition, image reasoning, captioning and answering general questions about an image as their main uses, with a 128K context window. One important detail: for tasks that combine images and text, English is the only officially supported language.

What Llama 3.2 is good at

• On-device AI. The 1B and 3B models are small enough for phones and lightweight laptops, so apps can work offline and keep data on the device.
• Fast, cheap text tasks. Summaries, rewrites, classification and short replies where a big model would be overkill.
• Image understanding in English. The vision models describe photos, read charts and documents, and answer questions about pictures.
• Experiments and learning. The small models are a popular first step for people trying local AI or fine-tuning for the first time.

Eight languages are officially supported for text: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.

Limits to know about

• Small means limited. The 1B and 3B models know far less than larger models and make more mistakes on facts and reasoning. Meta itself notes they are meant for highly constrained environments such as mobile devices.
• Older knowledge. Training data ends in December 2023.
• Vision is English-only for image-plus-text tasks, and the vision models do not generate images.
• EU restriction for the vision models. Meta's acceptable use policy for Llama 3.2 says the license rights for its multimodal models are not granted to individuals living in, or companies based in, the European Union. End users of a product that already includes these models are not affected. The 1B and 3B text models are not multimodal.
• Superseded for images. Llama 4 Scout and Maverick read images natively and support more languages.

How to get Llama 3.2 officially

Meta publishes all Llama 3.2 models on Hugging Face under the meta-llama organization, for example meta-llama/Llama-3.2-3B-Instruct and meta-llama/Llama-3.2-11B-Vision-Instruct. Accept the license on the model page to unlock the files. The vision model card is available here.

For local use, Ollama offers the small models: ollama run llama3.2 downloads the 3B version (about 2.0 GB) and ollama run llama3.2:1b the 1B version (about 1.3 GB). LM Studio and llama.cpp also run them; see run Llama locally.

Through cloud APIs, OpenRouter lists Llama 3.2 1B at $0.027 input and $0.201 output per million tokens, and Llama 3.2 3B at $0.0509 input and $0.335 output per million tokens (OpenRouter list price, Oct 2026). Our API guide shows how to call Llama models.

The license terms are the same in spirit as other Llama versions: "Built with Llama", the acceptable use policy, and the 700 million monthly active user rule. Our license page explains them.

Privacy notes

The main privacy benefit of Llama 3.2 is that the small models can run entirely on your device, so nothing leaves your phone or computer. If you use them through a cloud provider instead, the provider's privacy terms apply.

How Llama 3.2 compares

Against Llama 3.1 8B, the 3B model is smaller and lighter but less capable. Against Llama 4, the 3.2 vision models are older, support fewer languages for images and carry the same EU clause, while Llama 4 handles images in its main models. If you just want to ask about a photo, our chat with Llama 4 is the simplest option. The full sequence of releases is on Llama history.

FAQ

Can I use Llama 3.2 in the Llama.Online chat?

No, Llama 3.2 is not in our chat. Use Llama 4 Maverick or Scout for images and Llama 3.1 8B for fast text answers.

Can Llama 3.2 run on a phone?

The 1B and 3B models were designed for that, especially the quantized versions Meta released for mobile processors. You need an app that supports local models.

Which Llama 3.2 models can see images?

Only the 11B and 90B Vision models. The 1B and 3B models are text only.

Can people in the EU use Llama 3.2?

The text models, yes, under the license. For the vision models, Meta's policy does not grant the license rights to people and companies based in the EU, but end users of products built with them are not affected. This is general information, not legal advice.