Introducing Llama 3.2
Introducing
Llama 3.2
The open-source AI model you can fine-tune, distill and deploy anywhere is now available in more versions. Choose from 1B, 3B, 11B or 90B, or continue building with Llama 3.1
See how Llama is the leading open source model family
Latest models
Llama 3.2 is a collection of large language models (LLMs) pretrained and fine-tuned in 1B and 3B sizes that are multilingual text only, and 11B and 90B sizes that take both text and image inputs and output text.Start building
Lightweight
1B and 3B
Our lightweight and most efficient models you can run everywhere on mobile and on edge devices.Download models
Multimodal
11B and 90B
Our open multimodal models that are flexible and can reason on high resolution images.Download models
Download our flagship foundation 405B model
Do more with Llama 3.2
Develop highly performative and efficient applications from our latest release.
On-Device
Multimodal
Llama Stack
On-device
On-device
Use our 1B or 3B models for on device applications such as summarizing a discussion from your phone or calling on-device tools like calendar.
Multimodal
Multimodal
Use our 11B or 90B models for image use cases such as transforming an existing image into something new or getting more information from an image of your surroundings.
Llama Stack
On-device
Use our 1B or 3B models for on device applications such as summarizing a discussion from your phone or calling on-device tools like calendar.
On-device
Multimodal
Llama Stack
Use our 1B or 3B models for on-device applications such as summarizing a discussion from your phone or calling on-device tools like calendar.
Learn more
Llama Stack: a streamlined developer experience
Build faster, deploy anywhere and get the most out of the latest Llama models on day 1.
For developers
Best practices included
Optimized support for agentic tool calling, safety guardrails, inference and much more, significantly lowering development costs.Develop in your preferred language
Choose from python, node, kotlin, and swift programming languages to quickly build
your applications.
your applications.
Choose from python, node, kotlin, and swift programming languages to quickly build your applications.
Develop & deploy anywhere
With a common API, choose any distribution and deploy on-prem, locally hosted, or even on-device at the edge.
For partners & distributors
A standard API
Requires fewer model level changes across versions accelerating time to market for new models and lowering engineering investment.Interoperability with the ecosystem
Leverage the fast moving Llama ecosystem by building on a common API and incorporate new components faster.Support for agentic components
Llama Stack releases natively support tool calling, safety guardrails, retrieval augmented generation, an inference loop and other agentic functionality.
Llama Stack releases natively support tool calling, safety guardrails, retrieval augmented generation, an inference loop and other agentic functionality.
For developers
For developers
Best practices included
Optimized support for agentic tool calling, safety guardrails, inference and much more, significantly lowering development costs.Develop in your preferred language
Choose from python, node, kotlin, and swift programming languages to quickly build your applications.Develop & deploy anywhere
With a common API, choose any distribution and deploy on-prem, locally hosted, or even on-device at the edge.For Partners & Distributors
For Partners & Distributors
A standard API
Requires fewer model level changes across versions accelerating time to market for new models and lowering engineering investment.Interoperability with the ecosystem
Leverage the fast moving Llama ecosystem by building on a common API and incorporate new components faster.Support for agentic components
Llama Stack releases natively support tool calling, safety guardrails, retrieval augmented generation, an inference loop and other agentic functionality.Model evaluations
We evaluated performance on over 150 benchmark datasets that span a wide range of languages. For the vision LLMs, we evaluated performance on benchmarks for image understanding and visual reasoning. In addition, we performed extensive human evaluations that compare Llama 3.2 with competing models in real-world scenarios.
Instruction-tuned benchmarks
Lightweight instruction-tuned benchmarks
Vision instruction-tuned benchmarks
Category
Benchmark
General
MMLU(5-shot)
Open-rewrite eval(0-shot, rougeL)
TLDR9+(test, 1-shot, rougeL)
IFEval
Math
GSM8K(0-shot, CoT)
MATH(0-shot, CoT)
Reasoning
ARC Challenge(0-shot)
GPQA(0-shot)
Hellaswag(0-shot)
Tool use
BFCL V2
Nexus
Long context
InfiniteBench/En.MC(128k)
InfiniteBench/En.QA(128k)
NIH/Multi-needle
Multilingual
MGSM(0-shot, CoT)
Llama 3.2 1B
49.3
41.6
16.8
59.5
44.4
30.6
59.4
27.2
41.2
25.7
13.5
38.0
20.3
75.0
24.5
Llama 3.2 3B
63.4
40.1
19.0
77.4
77.7
48.0
78.6
32.8
69.8
67.0
34.3
63.3
19.8
84.7
58.2
Gemma 2 2B IT(5-shot)
57.8
31.2
13.9
61.9
62.5
23.8
76.7
27.5
61.1
27.4
21.0
-
-
-
40.2
Phi-3.5 - Mini IT(5-shot)
69.0
34.5
12.8
59.2
86.2
44.2
87.4
31.9
81.4
58.4
26.1
39.2
11.3
52.7
49.8
Lightweight
Vision
Category
Benchmark
General
MMLU(5-shot)
Open-rewrite eval(0-shot, rougeL)
TLDR9+(test, 1-shot, rougeL)
IFEval
Math
GSM8K(0-shot, CoT)
MATH(0-shot, CoT)
Reasoning
ARC Challenge(0-shot)
GPQA(0-shot)
Hellaswag(0-shot)
Tool use
BFCL V2
Nexus
Long context
InfiniteBench/En.MC(128k)
InfiniteBench/En.QA(128k)
NIH/Multi-needle
Multilingual
MGSM(0-shot, CoT)
Llama 3.2 1B
49.3
41.6
16.8
59.5
44.4
30.6
59.4
27.2
41.2
25.7
13.5
38.0
20.3
75.0
24.5
Llama 3.2 3B
63.4
40.1
19.0
77.4
77.7
48.0
78.6
32.8
69.8
67.0
34.3
63.3
19.8
84.7
58.2
Gemma 2 2B IT(5-shot)
57.8
31.2
13.9
61.9
62.5
23.8
76.7
27.5
61.1
27.4
21.0
-
-
-
40.2
Phi-3.5 - Mini IT(5-shot)
69.0
34.5
12.8
59.2
86.2
44.2
87.4
31.9
81.4