Explore the features of the new Llama 3.2 model, from image classification, vision reasoning to tool use.
Beginner1h29m
Introducing Multimodal Llama 3.2
Instructor: Amit Sangani
Earn an accomplishment with PRO · Updated Oct 7, 2025
- Beginner
- 1h29m
- 9 Video Lessons
- 6 Code Examples
- 1 Graded Assignment PRO
- Earn an accomplishment with PRO
- Updated Oct 7, 2025
- Instructor: Amit Sangani
Meta- Learn more aboutMembership PRO Plan
What you'll learn
Learn the details of Llama 3.2 prompting, tokenization, built-in and custom tool calling.
Gain knowledge of the Llama stack, which is a standardized interface for building AI applications.
Course recap
PROThis course covers Meta's Llama 3.2 model family, focusing on its new multimodal vision capabilities and tool calling features. Taught by Amit Sangani from Meta's Llama team, it provides hands-on experience with the 11B and 90B vision models for image reasoning tasks, the expanded prompt format with special tokens, tokenization improvements, and both built-in and custom tool calling.
Concept map
Key concepts
- Multimodal ModelsModels that process both text and images, enabled in Llama 3.2 through cross-attention layers between a pre-trained image encoder and text model
- Tool CallingLLM capability to invoke external functions (search, Wolfram Alpha, code interpreter) based on user queries
- TokenizationBreaking text into tokens using tiktoken with 128K vocabulary, 4x larger than Llama 2, improving encoding efficiency
- Prompt EngineeringStructured prompt format with special tokens (begin_of_text, start/end_header, end_of_turn) and roles (system, user, assistant, iPython)
Lesson highlights
- 1.Introduction — Overview of Llama 3.2 release with vision and lightweight models
- 2.Overview of Llama 3.2 — Model family details, training approach, benchmarks, and deployment options
- 3.Multimodal Prompting — Basic image reasoning with Llama 3.2 vision models (plant identification, dog breed recognition, car dashboard analysis)
- 4.Multimodal Use Cases — Eight advanced use cases including receipt analysis, diagram-to-code, chart-to-HTML, fridge analysis, and homework grading
- 5.Prompt Format — Detailed coverage of roles, special tokens, multi-turn chat, and message format
- 6.Tokenization — tiktoken tokenizer, 128K vocabulary, encoding efficiency, and how tokenization affects model understanding
- 7.Tool Calling — Built-in tools (Brave Search, Wolfram Alpha, code interpreter) and custom tool definition
- 8.Llama Stack — Meta's standardized APIs for inference, agents, memory, and safety
- 9.Conclusion — Summary and call to action
About this course
Join our new short course, Introducing Multimodal Llama 3.2, and learn from Amit Sangani, Senior Director of AI Partner Engineering at Meta, to learn all about the latest additions to the Llama models 3.1 and 3.2, from custom tool calling to multimodality and the new Llama stack.
Open models are a key building block of AI and a key enabler of AI research. With Meta’s family of open models, anyone can download, customize, fine-tune, or build new applications on top of them, allowing AI innovation. The Llama model family now ranges from 1B model parameters to its 405B foundation model, allowing for diverse use cases and applications.
In this course, you’ll learn about the new vision capabilities that Llama 3.2 brings to the Llama family. You’ll learn how to leverage this along with tool-calling, and Llama Stack, which is an open-source orchestration layer for building on top of the Llama family of models.
In detail, you’ll:
- Learn about the new models, how they were trained, their features, and how they fit into the Llama family.
- Understand how to do multimodal prompting with Llama and work on advanced image reasoning use cases such as understanding errors on a car dashboard, adding up the total of three restaurant receipts, grading written math homework, and many more.
- Learn different roles—system, user, assistant, ipython—in the Llama 3.1 and 3.2 family and the prompt format that identifies those roles.
- Understand how Llama uses the tiktoken tokenizer, and how it has expanded to a 128k vocabulary size that improves encoding efficiency and enables support for seven non-English languages.
- Learn how to prompt Llama to call both built-in and custom tools with examples for web search and solving math equations.
- Learn about ‘Llama Stack API’, which is a standardized interface for canonical toolchain components like fine-tuning or synthetic data generation to customize Llama models and build agentic applications.
Start building exciting applications on Llama!
Who should join?
Anyone who has basic Python knowledge and wants to learn to quickly build on Llama and Llama stack.
Course Outline
9 Lessons・6 Code Examples- IntroductionVideo・3m
- Overview of Llama 3.2Video・5m
- Multimodal PromptingVideo with Code Example・10m
- Multimodal Use CasesVideo with Code Example・14m
- Prompt FormatVideo with Code Example・12m
- TokenizationVideo with Code Example・7m
- Tool CallingVideo with Code Example・18m
- Llama StackVideo with Code Example・6m
- ConclusionVideo・1m
- Quiz
Graded・Quiz
・10m
Elevate your learning experience with Pro
Upgrade to Pro and gain unlimited accomplishments on your resume
Instructor
Course access is free for a limited time during the DeepLearning.AI learning platform beta!
Want to learn more about Generative AI?
Keep learning with updates on curated AI news, courses, and events, as well as Andrew’s thoughts from DeepLearning.AI!