SPDX Identifier: Llama-2 (unofficial)
- Meta’s proprietary license (not OSI-approved open source)
- Scale restriction: Services with 700 million or more monthly active users (MAU) require a separate license
- Commercial use: Allowed (within the scale restriction)
- Notice obligation: Retain the license information
- Use restrictions: Use for illegal or harmful purposes is prohibited
- Derivative models: State that they are based on Llama 2, and apply the same restrictions
To use Llama 2 in a product/service with 700 million or more monthly active users (MAU), you must obtain a separate license from Meta.
- Fewer than 700 million: Free to use
- 700 million or more: A license request to Meta is required
Most companies and services have fewer than 700 million users, so they can use it freely.
When developing a Llama 2-based service, please consult the OSPO.
What Is the Llama 2 Community License?
The Llama 2 Community License is a proprietary license that Meta released in 2023 together with the Llama 2 language model.
Llama Model Lineage
| Version | Release Date | License | Characteristics |
|---|---|---|---|
| Llama 1 | 2023.02 | Research only | Commercial use not allowed |
| Llama 2 | 2023.07 | Llama 2 Community | Commercial use allowed (scale restriction) |
| Llama 3 | 2024.04 | Llama 3 Community | Similar to Llama 2 (updated) |
Why a “Community License”?
Meta describes this license as follows:
- Encouraging community innovation: Most developers and companies can use it freely
- Checking large corporations: Restricting exclusive use by Big Tech
- Responsible AI: Preventing harmful use
700 Million Scale Restriction
Definition of MAU (Monthly Active Users)
Monthly active users: the number of unique users who used the product/service in the previous calendar month
Examples:
Scenario 1: T phone app (MAU 5 million)
Since the MAU is fewer than 700 million, it can be used freely.
Scenario 2: Facebook (MAU 3 billion)
Since the MAU is 700 million or more, a separate license must be requested from Meta.
Services Exceeding the 700 Million Threshold
Services with 700 million or more MAU worldwide:
- Facebook, Instagram, WhatsApp (Meta)
- YouTube (Google)
- TikTok
Most companies are not affected
When the Scale Restriction Applies
It applies continuously from the time of license agreement and after use begins.
Important: What happens if you start below 700 million but later exceed it? At the point of exceeding it, you must notify Meta and negotiate a separate license.
Permitted Items (MAU Fewer Than 700 Million)
What You Can Do Freely
Commercial use
- Providing paid services
- Selling APIs
- Integrating into products
Model modification
- Fine-tuning
- Quantization (4-bit, 8-bit)
- Model merging
Model distribution
- Releasing derivative models
- Uploading to Hugging Face, etc.
- Including in open source projects
Use of generated outputs
- Commercial use of AI-generated text
Prohibited Items (Acceptable Use Policy)
Llama 2 cannot be used for the following purposes:
1. Illegal Activities
- Supporting criminal acts
- Trading illegal goods
- Terrorist activities
2. Child Protection
- Child abuse content
- Child grooming
3. Discrimination and Hate
- Generating discriminatory content
- Hate speech
- Harassment
4. Violence and Danger
- Inciting violence
- Encouraging self-harm or suicide
- Weapons manufacturing
5. Privacy Violations
- Unauthorized collection of personal information
- Stalking, surveillance
6. Disinformation
- Intentional fake news
- Fraudulent content
7. High-Risk Domains (Sole Decision-Making)
- Medical diagnosis
- Legal advice
- Financial advice
- Emergency services
Prohibition on Training Competing LLMs with Llama 2
As a distinctive clause, you cannot use Llama 2 to train an LLM that competes with Meta.
Prohibited examples:
- Training a new LLM with the outputs of Llama 2
- Distillation using Llama 2 as a teacher model
Permitted examples:
- Fine-tuning Llama 2 itself
- Using the outputs of Llama 2 as a dataset (for a specific task)
Use Scenarios
Permitted Uses
1. Customer Service Chatbot
Scenario: A telecom customer consultation chatbot
MAU: 10 million
Judgment: Allowed (fewer than 700 million)
2. B2B AI Solution
Scenario: An enterprise document summarization solution
MAU: 100 thousand
Judgment: Allowed
3. Mobile App AI Feature
Scenario: A photo caption generation app
MAU: 5 million
Judgment: Allowed
Restricted Uses
1. Large-Scale Social Media
Scenario: Integration into social media with 1 billion MAU
Judgment: A Meta license is required
2. Training a Competing LLM
Scenario: Training a new LLM with Llama 2 outputs
Judgment: Prohibited
Requires Review
1. Global Service Expansion
Scenario: Currently 50 million MAU, with a future target of 1 billion
Judgment: Consultation with Meta is required before reaching 700 million
Licensing of Derivative Models
Mandatory Requirements
Attribution
- Indicate “Based on Llama 2”
- State it in the model card
License propagation
- Apply the same Acceptable Use Policy
- Maintain the 700 million scale restriction
Restriction on Meta trademark use
- Unauthorized use of the “Llama” trademark is prohibited
- The expression “Built with Llama 2” is OK
Derivative Model Examples
Permitted names:
- “MyModel (based on Llama 2)”
- “CustomLLM-Llama2-7B”
Prohibited names:
- “Llama 2 Pro”
- “Meta Llama Custom”
Llama 3 Update
Llama 3, released in April 2024, uses an updated license:
- Similar basic structure
- Some improvements to the use restriction clauses
- The 700 million scale restriction is maintained
The same guide applies when using Llama 3.