Llama 3.2 models are available for download on llama.com and Hugging Face, along with Meta’s broad ecosystem of partner platforms such as Google Cloud, IBM, Microsoft Azure, and NVIDIA.
Meta has introduced Llama 3.2, its open-source model with both image and text processing abilities. The Llama 3.2 model comes with small and medium-sized vision LLMs (11B and 90B) and lightweight, text-only models (1B and 3B) that support select edge and mobile devices.
Llama 3.2 (11B and 90B) to Support Image Reasoning
The two largest models in the Llama 3.2 collection, 11B and 90B, support image reasoning use cases such as document-level understanding, including charts and graphs, image captioning, and visual grounding tasks such as directionally pinpointing objects in images using natural language descriptions.
For example, a person can ask about the month’s best sales of their small business and Llama 3.2 can answer by reasoning using an available graph. The model can even respond with a map for questions like when a hike might become steeper.
The 11B and 90B models can also close the gap between vision and language by analyzing images, interpreting the scene, and generating a sentence or two as a caption to tell the story.
Llama 3.2 (1B and 3B) Capable with Text Generation
With the tool calling abilities and generating text in multiple languages, the lightweight 1B and 3B models are quite powerful. These models enable developers to create agentic, tailored, on-device applications with strong privacy that do not allow data to leave the device.
Such an application, for example, could help summarize the last 10 messages received, extract action items, and leverage tool calling to directly send calendar invites for follow-up meetings.
Running these models locally offers two major advantages.
- First, since processing is done locally, prompts and responses may appear immediate.
- Second, running models locally maintains privacy and enhances the overall application’s privacy by not sending data such as calendar and messaging information to the cloud.
The application can explicitly control which queries remain on the device and which may need to be processed by a larger model in the cloud as processing is done locally.
Models’ Performance
Meta’s evaluation suggests that Llama 3.2 vision models are competitive with leading foundation models, Claude 3 Haiku and GPT4o-mini on image recognition and various visual understanding tasks.
According to the company, the 3B model outperformed the Gemma 2 2.6B and Phi 3.5-mini models at certain tasks like prompt rewriting, summarization, and instruction following. Further, the company said 1B is competitive with Gemma.
Meta used more than 150 benchmark datasets in a variety of languages to assess the models’ performance. The company evaluated the vision LLMs’ performance on benchmark for visual reasoning and image understanding.
Training
The 11B and 90B models, as the first Llama models to support vision tasks, required an entirely new model architecture that supports image reasoning.
Meta trained a set of adapter weights that integrate the previously trained image encoder into the previously trained language model in order to add image input support.
The training pipeline encompasses multiple stages, starting from pre-trained Llama 3.1 text models.
The end result is a set of models that are able to understand and reason a combination of image and text prompts. This is another step toward more richer agentic capabilities for Llama models.
Llama Guard 3 11B – System level Safety
Meta has also released Llama Guard 3 11B Vision designed to support Llama 3.2’s new image understanding capability and filter text+image input prompts or text output responses to these prompts.
As the company introduced 1B and 3B Llama models to use in more constrained environments like on-device, it also optimized Llama Guard to significantly reduce its deployment cost.
Based on the Llama 3.2 1B model, Llama Guard 3 1B has been quantized and pruned, reducing its size from 2,858 MB to 438 MB, making deployment more efficient than ever.
Models’ Availability
Llama 3.2 models are available for download on llama.com and Hugging Face. The models are also available for immediate development on Meta’s broad ecosystem of partner platforms, including AWS, AMD, Google Cloud, Groq, Databricks, Dell, IBM, Microsoft Azure, NVIDIA, Intel, Oracle Cloud, Snowflake, and more.
According to Meta, Llama has achieved 10x growth, with over 350 million downloads to date, leading AI innovation.
Also Read:
This is a great example of how AI can be used to save time, just like James Jernigan always says. He’s a big advocate for using technology to work smarter, not harder. I’ve learned so much from him.