Llama 3.2
Small text models and vision models from Meta for edge devices and image tasks
Llama 3.2 is a set of four open models released by Meta on 25 September 2024: two lightweight text-only models (1B and 3B) and two vision models (11B and 90B). It matters because it brings Llama to phones and edge hardware, and because it is the first Llama line that can read images.
- Four sizes: 1B and 3B text-only, 11B and 90B with image input
- 1B and 3B handle a 128K token context window
- First Llama models able to take images as well as text
- Vision models keep Llama 3.1 text skills by freezing language weights
- Announced at Connect on 25 September 2024
What it is
Llama 3.2 is a family of open-weight language models published by Meta in September 2024. It has four sizes. Two of them, at 1 billion and 3 billion parameters, handle text only and are built to run on phones and other edge hardware. The other two, at 11 billion and 90 billion parameters, accept images alongside text. Each size ships in both a pre-trained and an instruction-tuned form. Meta positions the small pair for summarising, rewriting and following instructions locally, and the larger pair for reading charts, captions and maps.
How it came about
Meta announced the release at its Connect event on 25 September 2024, with Mark Zuckerberg presenting it. It arrived two months after the Llama 3.1 herd, which included the 405B model that Meta described as the first open frontier-level system. The company says Llama had grown tenfold over the past year and had become a standard for responsible innovation. Meta also put out the first official Llama Stack distributions at the same time, working with more than 25 firms.
Inside the model
The vision models needed a fresh architecture because earlier Llama versions could not take images. Meta trained adapter weights that connect a pre-trained image encoder to a pre-trained language model through a run of cross-attention layers. Training used text-and-image pairs, first large volumes of noisy data and then smaller, higher quality in-domain data. The image encoder was updated during this stage but the language model parameters were left untouched, so text skills stayed intact and the models can swap in for Llama 3.1. Safety data was added too.
What it can do
Meta's own evaluations cover more than 150 benchmark datasets across many languages, plus image understanding and visual reasoning tests. The company reports the vision models are competitive with Claude 3 Haiku and GPT4o-mini on image recognition and visual understanding. It says the 3B model beats Gemma 2 2.6B and Phi 3.5-mini on instruction following, summarisation, prompt rewriting and tool use, while 1B is on a par with Gemma. The small models handle multilingual generation and tool calling. Vision tasks include chart and graph reading, captioning and pointing at objects from a description.
How to use it
Weights are downloadable from llama.com and Hugging Face, and the models run on partner services from launch. Both pre-trained and aligned versions can be fine-tuned with torchtune and deployed locally with torchchat. Llama Stack distributions cover single-node, on-premises, cloud and on-device use, with PyTorch ExecuTorch for devices and Ollama for single-node setups. Availability of the multimodal models in Europe is governed by the acceptable use policy.
Safety and testing
Meta says safety mitigation data was mixed into fine-tuning so the models keep a high level of safety while staying helpful. Llama Stack ships with integrated safety for retrieval and tooling applications. Details on multimodal availability in Europe sit in the acceptable use policy.
Reception and debate
Meta argues that openness helps developers, itself and the wider world, and that Llama leads on openness, modifiability and cost efficiency while staying competitive with closed models. The company frames the release as a response to developers who lacked the compute needed for Llama 3.1.
Why it matters
The release extends Llama to hardware that could not run earlier versions, with day-one support on Qualcomm and MediaTek chips and optimisation for Arm processors. Vision support is a first for the family, and Meta describes it as a step towards richer agentic abilities. Llama Stack is meant to simplify how developers deploy Llama across different environments, including retrieval-augmented generation and tool-using apps.
Questions people ask about Llama 3.2
What sizes does Llama 3.2 come in?
There are four: 1B and 3B parameter models that work on text alone, plus 11B and 90B parameter models that accept images alongside text. Every size is available both pre-trained and instruction-tuned.
Can it run on a phone?
The 1B and 3B models are designed for edge and mobile hardware. Meta says they work on Qualcomm and MediaTek chips from day one and are tuned for Arm processors. The larger vision models need more resources.
How do the vision models handle images?
An image encoder feeds into the language model through cross-attention adapter layers trained on image and text pairs. The language model itself was not updated, so text abilities remain and the models can replace Llama 3.1 directly.
How do I get the models?
Weights are free to download from llama.com and Hugging Face. You can fine-tune them with torchtune, run them locally with torchchat, or use Llama Stack distributions for on-device, single-node, on-premises and cloud deployment.
What can the vision models do?
Meta highlights reading charts and graphs, captioning images, and locating objects from a spoken description. The company gives examples such as finding a best sales month from a graph or judging how steep a trail is from a map.
Good reading and sources
There is no English Wikipedia article on Llama 3.2 yet, so the facts on this page come from the sources listed above: ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.