The Llama 3.2 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction-tuned generative models in 1B and 3B sizes (text in/text out). The Llama 3.2 instruction-tuned text only models are optimized for multilingual dialogue use cases, including agentic retrieval and summarization tasks. They outperform many of the available open source and closed chat models on common industry benchmarks.
Model Developer: Meta
Model Architecture: Llama 3.2 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety.
Training Data
Params
Input modalities
Output modalities
Context Length
GQA
Shared Embeddings
Token count
Knowledge cutoff
Llama 3.2 (text only)
A new mix of publicly available online data.
1B (1.23B)
Multilingual Text
Multilingual Text and code
128k
Yes
Yes
Up to 9T tokens
December 2023
3B (3.21B)
Multilingual Text
Multilingual Text and code
Llama 3.2 Quantized (text only)
A new mix of publicly available online data.
1B (1.23B)
Multilingual Text
Multilingual Text and code
8k
Yes
Yes
Up to 9T tokens
December 2023
3B (3.21B)
Multilingual Text
Multilingual Text and code
Supported Languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai are officially supported. Llama 3.2 has been trained on a broader collection of languages than these 8 supported languages. Developers may fine-tune Llama 3.2 models for languages beyond these supported languages, provided they comply with the Llama 3.2 Community License and the Acceptable Use Policy. Developers are always expected to ensure that their deployments, including those that involve additional languages, are completed safely and responsibly.
Llama 3.2 Model Family: Token counts refer to pretraining data only. All model versions use Grouped-Query Attention (GQA) for improved inference scalability.
Model Release Date: Sept 25, 2024
Status: This is a static model trained on an offline dataset. Future versions may be released that improve model capabilities and safety.
License: Use of Llama 3.2 is governed by the Llama 3.2 Community License (a custom, commercial license agreement).
Feedback: Instructions on how to provide feedback or comments on the model can be found in the Llama Models README. For more technical information about generation parameters and recipes for how to use Llama 3.2 in applications, please go here.
Intended Use
Intended Use Cases: Llama 3.2 is intended for commercial and research use in multiple languages. Instruction tuned text only models are intended for assistant-like chat and agentic applications like knowledge retrieval and summarization, mobile AI powered writing assistants and query and prompt rewriting. Pretrained models can be adapted for a variety of additional natural language generation tasks. Similarly, quantized models can be adapted for a variety of on-device use-cases with limited compute resources.
Out of Scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in any other way that is prohibited by the Acceptable Use Policy and Llama 3.2 Community License. Use in languages beyond those explicitly referenced as supported in this model card.
How to use
This repository contains two versions of Llama-3.2-3B-Instruct, for use with transformers and with the original llama codebase.
Use with transformers
Starting with transformers >= 4.43.0 onward, you can run conversational inference using the Transformers pipeline abstraction or by leveraging the Auto classes with the generate() function.
Make sure to update your transformers installation via pip install --upgrade transformers.
python
1import torch
2from transformers import pipeline
34model_id ="meta-llama/Llama-3.2-3B-Instruct"5pipe = pipeline(6"text-generation",7 model=model_id,8 torch_dtype=torch.bfloat16,9 device_map="auto",10)11messages =[12{"role":"system","content":"You are a pirate chatbot who always responds in pirate speak!"},13{"role":"user","content":"Who are you?"},14]15outputs = pipe(16 messages,17 max_new_tokens=256,18)19print(outputs[0]["generated_text"][-1])
Note: You can also find detailed recipes on how to use the model locally, with torch.compile(), assisted generations, quantised and more at huggingface-llama-recipes
Training Factors: We used custom training libraries, Meta's custom built GPU cluster, and production infrastructure for pretraining. Fine-tuning, quantization, annotation, and evaluation were also performed on production infrastructure.
Training Energy Use: Training utilized a cumulative of 916k GPU hours of computation on H100-80GB (TDP of 700W) type hardware, per the table below. Training time is the total GPU time required for training each model and power consumption is the peak power capacity per GPU device used, adjusted for power usage efficiency.
Training Greenhouse Gas Emissions: Estimated total location-based greenhouse gas emissions were 240 tons CO2eq for training. Since 2020, Meta has maintained net zero greenhouse gas emissions in its global operations and matched 100% of its electricity use with renewable energy; therefore, the total market-based greenhouse gas emissions for training were 0 tons CO2eq.
Training Time (GPU hours)
Logit Generation Time (GPU Hours)
Training Power Consumption (W)
Training Location-Based Greenhouse Gas Emissions (tons CO2eq)
Training Market-Based Greenhouse Gas Emissions (tons CO2eq)
Llama 3.2 1B
370k
-
700
107
0
Llama 3.2 3B
460k
-
700
133
0
Llama 3.2 1B SpinQuant
1.7
0
700
Negligible**
0
Llama 3.2 3B SpinQuant
2.4
0
700
Negligible**
0
Llama 3.2 1B QLora
1.3k
0
700
0.381
0
Llama 3.2 3B QLora
1.6k
0
700
0.461
0
Total
833k
86k
240
0
** The location-based CO2e emissions of Llama 3.2 1B SpinQuant and Llama 3.2 3B SpinQuant are less than 0.001 metric tonnes each. This is due to the minimal training GPU hours that are required.
The methodology used to determine training energy use and greenhouse gas emissions can be found here. Since Meta is openly releasing these models, the training energy use and greenhouse gas emissions will not be incurred by others.
Training Data
Overview: Llama 3.2 was pretrained on up to 9 trillion tokens of data from publicly available sources. For the 1B and 3B Llama 3.2 models, we incorporated logits from the Llama 3.1 8B and 70B models into the pretraining stage of the model development, where outputs (logits) from these larger models were used as token-level targets. Knowledge distillation was used after pruning to recover performance. In post-training we used a similar recipe as Llama 3.1 and produced final chat models by doing several rounds of alignment on top of the pre-trained model. Each round involved Supervised Fine-Tuning (SFT), Rejection Sampling (RS), and Direct Preference Optimization (DPO).
Data Freshness: The pretraining data has a cutoff of December 2023.
Quantization
Quantization Scheme
We designed the current quantization scheme with the PyTorch’s ExecuTorch inference framework and Arm CPU backend in mind, taking into account metrics including model quality, prefill/decoding speed, and memory footprint. Our quantization scheme involves three parts:
All linear layers in all transformer blocks are quantized to a 4-bit groupwise scheme (with a group size of 32) for weights and 8-bit per-token dynamic quantization for activations.
The classification layer is quantized to 8-bit per-channel for weight and 8-bit per token dynamic quantization for activation.
Similar to classification layer, an 8-bit per channel quantization is used for embedding layer.
Quantization-Aware Training and LoRA
The quantization-aware training (QAT) with low-rank adaptation (LoRA) models went through only post-training stages, using the same data as the full precision models. To initialize QAT, we utilize BF16 Llama 3.2 model checkpoints obtained after supervised fine-tuning (SFT) and perform an additional full round of SFT training with QAT. We then freeze the backbone of the QAT model and perform another round of SFT with LoRA adaptors applied to all layers within the transformer block. Meanwhile, the LoRA adaptors' weights and activations are maintained in BF16. Because our approach is similar to QLoRA of Dettmers et al., (2023) (i.e., quantization followed by LoRA adapters), we refer this method as QLoRA. Finally, we fine-tune the resulting model (both backbone and LoRA adaptors) using direct preference optimization (DPO).
SpinQuant
SpinQuant was applied, together with generative post-training quantization (GPTQ). For the SpinQuant rotation matrix fine-tuning, we optimized for 100 iterations, using 800 samples with sequence-length 2048 from the WikiText 2 dataset. For GPTQ, we used 128 samples from the same dataset with the same sequence-length.
Benchmarks - English Text
In this section, we report the results for Llama 3.2 models on standard automatic benchmarks. For all these evaluations, we used our internal evaluations library.
Base Pretrained Models
Category
Benchmark
# Shots
Metric
Llama 3.2 1B
Llama 3.2 3B
Llama 3.1 8B
General
MMLU
5
macro_avg/acc_char
32.2
58
66.7
AGIEval English
3-5
average/acc_char
23.3
39.2
47.8
ARC-Challenge
25
acc_char
32.8
69.1
79.7
Reading comprehension
SQuAD
1
em
49.2
67.7
77
QuAC (F1)
1
f1
37.9
42.9
44.9
DROP (F1)
3
f1
28.0
45.2
59.5
Long Context
Needle in Haystack
0
em
96.8
1
1
Instruction Tuned Models
Capability
Benchmark
# Shots
Metric
Llama 3.2 1B bf16
Llama 3.2 1B Vanilla PTQ**
Llama 3.2 1B Spin Quant
Llama 3.2 1B QLoRA
Llama 3.2 3B bf16
Llama 3.2 3B Vanilla PTQ**
Llama 3.2 3B Spin Quant
Llama 3.2 3B QLoRA
Llama 3.1 8B
General
MMLU
5
macro_avg/acc
49.3
43.3
47.3
49.0
63.4
60.5
62
62.4
69.4
Re-writing
Open-rewrite eval
0
micro_avg/rougeL
41.6
39.2
40.9
41.2
40.1
40.3
40.8
40.7
40.9
Summarization
TLDR9+ (test)
1
rougeL
16.8
14.9
16.7
16.8
19.0
19.1
19.2
19.1
17.2
Instruction following
IFEval
0
Avg(Prompt/Instruction acc Loose/Strict)
59.5
51.5
58.4
55.6
77.4
73.9
73.5
75.9
80.4
Math
GSM8K (CoT)
8
em_maj1@1
44.4
33.1
40.6
46.5
77.7
72.9
75.7
77.9
84.5
MATH (CoT)
0
final_em
30.6
20.5
25.3
31.0
48.0
44.2
45.3
49.2
51.9
Reasoning
ARC-C
0
acc
59.4
54.3
57
60.7
78.6
75.6
77.6
77.6
83.4
GPQA
0
acc
27.2
25.9
26.3
25.9
32.8
32.8
31.7
33.9
32.8
Hellaswag
0
acc
41.2
38.1
41.3
41.5
69.8
66.3
68
66.3
78.7
Tool Use
BFCL V2
0
acc
25.7
14.3
15.9
23.7
67.0
53.4
60.1
63.5
67.1
Nexus
0
macro_avg/acc
13.5
5.2
9.6
12.5
34.3
32.4
31.5
30.1
38.5
Long Context
InfiniteBench/En.QA
0
longbook_qa/f1
20.3
N/A
N/A
N/A
19.8
N/A
N/A
N/A
27.3
InfiniteBench/En.MC
0
longbook_choice/acc
38.0
N/A
N/A
N/A
63.3
N/A
N/A
N/A
72.2
NIH/Multi-needle
0
recall
75.0
N/A
N/A
N/A
84.7
N/A
N/A
N/A
98.8
Multilingual
MGSM (CoT)
0
em
24.5
13.7
18.2
24.4
58.2
48.9
54.3
56.8
68.9
**for comparison purposes only. Model not released.
Multilingual Benchmarks
Category
Benchmark
Language
Llama 3.2 1B
Llama 3.2 1B Vanilla PTQ**
Llama 3.2 1B Spin Quant
Llama 3.2 1B QLoRA
Llama 3.2 3B
Llama 3.2 3B Vanilla PTQ**
Llama 3.2 3B Spin Quant
Llama 3.2 3B QLoRA
Llama 3.1 8B
General
MMLU (5-shot, macro_avg/acc)
Portuguese
39.8
34.9
38.9
40.2
54.5
50.9
53.3
53.4
62.1
Spanish
41.5
36.0
39.8
41.8
55.1
51.9
53.6
53.6
62.5
Italian
39.8
34.9
38.1
40.6
53.8
49.9
52.1
51.7
61.6
German
39.2
34.9
37.5
39.6
53.3
50.0
52.2
51.3
60.6
French
40.5
34.8
39.2
40.8
54.6
51.2
53.3
53.3
62.3
Hindi
33.5
30.0
32.1
34.0
43.3
40.4
42.0
42.1
50.9
Thai
34.7
31.2
32.4
34.9
44.5
41.3
44.0
42.2
50.3
**for comparison purposes only. Model not released.
Inference time
In the below table, we compare the performance metrics of different quantization methods (SpinQuant and QAT + LoRA) with the BF16 baseline. The evaluation was done using the ExecuTorch framework as the inference engine, with the ARM CPU as a backend using Android OnePlus 12 device.
Category
Decode (tokens/sec)
Time-to-first-token (sec)
Prefill (tokens/sec)
Model size (PTE file size in MB)
Memory size (RSS in MB)
1B BF16 (baseline)
19.2
1.0
60.3
2358
3,185
1B SpinQuant
50.2 (2.6x)
0.3 (-76.9%)
260.5 (4.3x)
1083 (-54.1%)
1,921 (-39.7%)
1B QLoRA
45.8 (2.4x)
0.3 (-76.0%)
252.0 (4.2x)
1127 (-52.2%)
2,255 (-29.2%)
3B BF16 (baseline)
7.6
3.0
21.2
6129
7,419
3B SpinQuant
19.7 (2.6x)
0.7 (-76.4%)
89.7 (4.2x)
2435 (-60.3%)
3,726 (-49.8%)
3B QLoRA
18.5 (2.4x)
0.7 (-76.1%)
88.8 (4.2x)
2529 (-58.7%)
4,060 (-45.3%)
(*) The performance measurement is done using an adb binary-based approach.
(**) It is measured on an Android OnePlus 12 device.
(***) Time-to-first-token (TTFT) is measured with prompt length=64
Footnote:
Decode (tokens/second) is for how quickly it keeps generating. Higher is better.
Time-to-first-token (TTFT for shorthand) is for how fast it generates the first token for a given prompt. Lower is better.
Prefill is the inverse of TTFT (aka 1/TTFT) in tokens/second. Higher is better
Model size - how big is the model, measured by, PTE file, a binary file format for ExecuTorch
RSS size - Memory usage in resident set size (RSS)
Responsibility & Safety
As part of our Responsible release approach, we followed a three-pronged strategy to managing trust & safety risks:
Enable developers to deploy helpful, safe and flexible experiences for their target audience and for the use cases supported by Llama
Protect developers against adversarial users aiming to exploit Llama capabilities to potentially cause harm
Provide protections for the community to help prevent the misuse of our models
Responsible Deployment
Approach: Llama is a foundational technology designed to be used in a variety of use cases. Examples on how Meta’s Llama models have been responsibly deployed can be found in our Community Stories webpage. Our approach is to build the most helpful models, enabling the world to benefit from the technology power, by aligning our model safety for generic use cases and addressing a standard set of harms. Developers are then in the driver’s seat to tailor safety for their use cases, defining their own policies and deploying the models with the necessary safeguards in their Llama systems. Llama 3.2 was developed following the best practices outlined in our Responsible Use Guide.
Llama 3.2 Instruct
Objective: Our main objectives for conducting safety fine-tuning are to provide the research community with a valuable resource for studying the robustness of safety fine-tuning, as well as to offer developers a readily available, safe, and powerful model for various applications to reduce the developer workload to deploy safe AI systems. We implemented the same set of safety mitigations as in Llama 3, and you can learn more about these in the Llama 3 paper.
Fine-Tuning Data: We employ a multi-faceted approach to data collection, combining human-generated data from our vendors with synthetic data to mitigate potential safety risks. We’ve developed many large language model (LLM)-based classifiers that enable us to thoughtfully select high-quality prompts and responses, enhancing data quality control.
Refusals and Tone: Building on the work we started with Llama 3, we put a great emphasis on model refusals to benign prompts as well as refusal tone. We included both borderline and adversarial prompts in our safety data strategy, and modified our safety data responses to follow tone guidelines.
Llama 3.2 Systems
Safety as a System: Large language models, including Llama 3.2, are not designed to be deployed in isolation but instead should be deployed as part of an overall AI system with additional safety guardrails as required. Developers are expected to deploy system safeguards when building agentic systems. Safeguards are key to achieve the right helpfulness-safety alignment as well as mitigating safety and security risks inherent to the system and any integration of the model or system with external tools. As part of our responsible release approach, we provide the community with safeguards that developers should deploy with Llama models or other LLMs, including Llama Guard, Prompt Guard and Code Shield. All our reference implementations demos contain these safeguards by default so developers can benefit from system-level safety out-of-the-box.
New Capabilities and Use Cases
Technological Advancement: Llama releases usually introduce new capabilities that require specific considerations in addition to the best practices that generally apply across all Generative AI use cases. For prior release capabilities also supported by Llama 3.2, see Llama 3.1 Model Card, as the same considerations apply here as well.
Constrained Environments: Llama 3.2 1B and 3B models are expected to be deployed in highly constrained environments, such as mobile devices. LLM Systems using smaller models will have a different alignment profile and safety/helpfulness tradeoff than more complex, larger systems. Developers should ensure the safety of their system meets the requirements of their use case. We recommend using lighter system safeguards for such use cases, like Llama Guard 3-1B or its mobile-optimized version.
Evaluations
Scaled Evaluations: We built dedicated, adversarial evaluation datasets and evaluated systems composed of Llama models and Purple Llama safeguards to filter input prompt and output response. It is important to evaluate applications in context, and we recommend building dedicated evaluation dataset for your use case.
Red Teaming: We conducted recurring red teaming exercises with the goal of discovering risks via adversarial prompting and we used the learnings to improve our benchmarks and safety tuning datasets. We partnered early with subject-matter experts in critical risk areas to understand the nature of these real-world harms and how such models may lead to unintended harm for society. Based on these conversations, we derived a set of adversarial goals for the red team to attempt to achieve, such as extracting harmful information or reprogramming the model to act in a potentially harmful capacity. The red team consisted of experts in cybersecurity, adversarial machine learning, responsible AI, and integrity in addition to multilingual content specialists with background in integrity issues in specific geographic markets.
Critical Risks
In addition to our safety work above, we took extra care on measuring and/or mitigating the following critical risk areas:
1. CBRNE (Chemical, Biological, Radiological, Nuclear, and Explosive Weapons): Llama 3.2 1B and 3B models are smaller and less capable derivatives of Llama 3.1. For Llama 3.1 70B and 405B, to assess risks related to proliferation of chemical and biological weapons, we performed uplift testing designed to assess whether use of Llama 3.1 models could meaningfully increase the capabilities of malicious actors to plan or carry out attacks using these types of weapons and have determined that such testing also applies to the smaller 1B and 3B models.
2. Child Safety: Child Safety risk assessments were conducted using a team of experts, to assess the model’s capability to produce outputs that could result in Child Safety risks and inform on any necessary and appropriate risk mitigations via fine tuning. We leveraged those expert red teaming sessions to expand the coverage of our evaluation benchmarks through Llama 3 model development. For Llama 3, we conducted new in-depth sessions using objective based methodologies to assess the model risks along multiple attack vectors including the additional languages Llama 3 is trained on. We also partnered with content specialists to perform red teaming exercises assessing potentially violating content while taking account of market specific nuances or experiences.
3. Cyber Attacks: For Llama 3.1 405B, our cyber attack uplift study investigated whether LLMs can enhance human capabilities in hacking tasks, both in terms of skill level and speed.
Our attack automation study focused on evaluating the capabilities of LLMs when used as autonomous agents in cyber offensive operations, specifically in the context of ransomware attacks. This evaluation was distinct from previous studies that considered LLMs as interactive assistants. The primary objective was to assess whether these models could effectively function as independent agents in executing complex cyber-attacks without human intervention. Because Llama 3.2’s 1B and 3B models are smaller and less capable models than Llama 3.1 405B, we broadly believe that the testing conducted for the 405B model also applies to Llama 3.2 models.
Community
Industry Partnerships: Generative AI safety requires expertise and tooling, and we believe in the strength of the open community to accelerate its progress. We are active members of open consortiums, including the AI Alliance, Partnership on AI and MLCommons, actively contributing to safety standardization and transparency. We encourage the community to adopt taxonomies like the MLCommons Proof of Concept evaluation to facilitate collaboration and transparency on safety and content evaluations. Our Purple Llama tools are open sourced for the community to use and widely distributed across ecosystem partners including cloud service providers. We encourage community contributions to our Github repository.
Grants: We also set up the Llama Impact Grants program to identify and support the most compelling applications of Meta’s Llama model for societal benefit across three categories: education, climate and open innovation. The 20 finalists from the hundreds of applications can be found here.
Reporting: Finally, we put in place a set of resources including an output reporting mechanism and bug bounty program to continuously improve the Llama technology with the help of the community.
Ethical Considerations and Limitations
Values: The core values of Llama 3.2 are openness, inclusivity and helpfulness. It is meant to serve everyone, and to work for a wide range of use cases. It is thus designed to be accessible to people across many different backgrounds, experiences and perspectives. Llama 3.2 addresses users and their needs as they are, without insertion unnecessary judgment or normativity, while reflecting the understanding that even content that may appear problematic in some cases can serve valuable purposes in others. It respects the dignity and autonomy of all users, especially in terms of the values of free thought and expression that power innovation and progress.
Testing: Llama 3.2 is a new technology, and like any new technology, there are risks associated with its use. Testing conducted to date has not covered, nor could it cover, all scenarios. For these reasons, as with all LLMs, Llama 3.2’s potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts. Therefore, before deploying any applications of Llama 3.2 models, developers should perform safety testing and tuning tailored to their specific applications of the model. Please refer to available resources including our Responsible Use Guide, Trust and Safety solutions, and other resources to learn more about responsible development.
