# Llama Meta's Llama model ecosystem encompasses Llama Guard 4, a natively multimodal 12B-parameter safeguard model designed for the Llama 4 and Llama 3 model lines, as well as broader deployment options and prompt formatting details for Llama models across various platforms. Status: provisional Last verified: 2026-08-28T10:06:29.000Z ## Evidence - (single_source) Llama Guard 4 is a natively multimodal safeguard model with 12 billion parameters that uses an early fusion transformer architecture with dense layers. - supports https://llama.com/docs/model-cards-and-prompt-formats/llama-guard-4 > Llama Guard 4 (12B) is our latest safeguard model with improved inference for detecting problematic prompts and responses. It is designed to work with the Llama 4 line of models, such as Llama 4 Scout and Llama 4 Maverick. ... Llama Guard 4 is a natively multimodal safeguard model. The model has 12 billion parameters in total and uses an early fusion transformer architecture with dense layers to keep the overall size small. The model can be run on a single GPU. Llama Guard 4 shares the same tokenizer and vision encoder as Llama 4 Scout and Maverick. ... Llama Guard 4 is also compatible with the Llama 3 line of models and can be used as a drop-in replacement for Llama Guard 3 8B and 11B for both text-only and multimodal applications. However, Llama Guard 3 1B still holds value for specific - (single_source) Llama Guard 4 can be run on a single GPU. - supports https://llama.com/docs/model-cards-and-prompt-formats/llama-guard-4 > Llama Guard 4 (12B) is our latest safeguard model with improved inference for detecting problematic prompts and responses. It is designed to work with the Llama 4 line of models, such as Llama 4 Scout and Llama 4 Maverick. ... Llama Guard 4 is a natively multimodal safeguard model. The model has 12 billion parameters in total and uses an early fusion transformer architecture with dense layers to keep the overall size small. The model can be run on a single GPU. Llama Guard 4 shares the same tokenizer and vision encoder as Llama 4 Scout and Maverick. ... Llama Guard 4 is also compatible with the Llama 3 line of models and can be used as a drop-in replacement for Llama Guard 3 8B and 11B for both text-only and multimodal applications. However, Llama Guard 3 1B still holds value for specific - (single_source) Meta Llama models can be used in contexts beyond Cloud Service Providers, including Linux, the Windows Subsystem for Linux (WSL), macOS, Jupyter notebooks, and mobile devices. - supports https://llama.com/docs/llama-everywhere > Although Meta Llama models are often hosted by Cloud Service Providers (CSP), Meta Llama can be used in other contexts as well, such as Linux, the Windows Subsystem for Linux (WSL), macOS, Jupyter notebooks, and even mobile devices. If you are interested in exploring these scenarios, we suggest that you check out the following resources: ... - Llama 3 on Your Local Computer, with Resources for Other Options - How to run Llama on your desktop using Windows, macOS, or Linux. Also, pointers to other ways to run Llama, either on premise or in the cloud ... - Llama Recipes QuickStart- Provides an introduction to Meta Llama using Jupyter notebooks and also demonstrates running Llama locally on macOS. ... - Llama.cpp- Uses the portability of C++ to enable inference with Llama models on a variety - (single_source) Llama 3 uses special tokens including an end-of-text token equivalent to the EOS token, on which generation ceases. - supports https://llama.com/docs/model-cards-and-prompt-formats/other-models > # Other Models ... - Llama 3 - Llama Guard 2 - Code Llama 70B - Llama Guard 1 - Code Llama - Llama 2 ... ### Llama 3 ... You can find details about this model in the model card. ... Special Tokens used with Llama 3 ... a particular message ... The possible roles can be: system, user, assistant. ... <|end_of_text|>: This is equivalent to the EOS token. On generating this token, Llama 3 will cease to generate more tokens. ... A prompt can optionally contain a single system message, or multiple alternating user and assistant messages, but always ends with the last user message followed by the assistant header. ... Llama 3 ... Note: Newlines (0x0A) are part of the prompt format, for clarity in the example, they have been represented as actual new lines. ... begin_of_text ... Llama 3 Instruct - (single_source) Llama.cpp uses the portability of C++ to enable inference with Llama models on a variety of platforms. - supports https://llama.com/docs/llama-everywhere > Although Meta Llama models are often hosted by Cloud Service Providers (CSP), Meta Llama can be used in other contexts as well, such as Linux, the Windows Subsystem for Linux (WSL), macOS, Jupyter notebooks, and even mobile devices. If you are interested in exploring these scenarios, we suggest that you check out the following resources: ... - Llama 3 on Your Local Computer, with Resources for Other Options - How to run Llama on your desktop using Windows, macOS, or Linux. Also, pointers to other ways to run Llama, either on premise or in the cloud ... - Llama Recipes QuickStart- Provides an introduction to Meta Llama using Jupyter notebooks and also demonstrates running Llama locally on macOS. ... - Llama.cpp- Uses the portability of C++ to enable inference with Llama models on a variety - (single_source) In Llama 3, the possible message roles are system, user, and assistant. - supports https://llama.com/docs/model-cards-and-prompt-formats/other-models > # Other Models ... - Llama 3 - Llama Guard 2 - Code Llama 70B - Llama Guard 1 - Code Llama - Llama 2 ... ### Llama 3 ... You can find details about this model in the model card. ... Special Tokens used with Llama 3 ... a particular message ... The possible roles can be: system, user, assistant. ... <|end_of_text|>: This is equivalent to the EOS token. On generating this token, Llama 3 will cease to generate more tokens. ... A prompt can optionally contain a single system message, or multiple alternating user and assistant messages, but always ends with the last user message followed by the assistant header. ... Llama 3 ... Note: Newlines (0x0A) are part of the prompt format, for clarity in the example, they have been represented as actual new lines. ... begin_of_text ... Llama 3 Instruct - (single_source) Llama Recipes QuickStart provides an introduction to Meta Llama using Jupyter notebooks and demonstrates running Llama locally on macOS. - supports https://llama.com/docs/llama-everywhere > Although Meta Llama models are often hosted by Cloud Service Providers (CSP), Meta Llama can be used in other contexts as well, such as Linux, the Windows Subsystem for Linux (WSL), macOS, Jupyter notebooks, and even mobile devices. If you are interested in exploring these scenarios, we suggest that you check out the following resources: ... - Llama 3 on Your Local Computer, with Resources for Other Options - How to run Llama on your desktop using Windows, macOS, or Linux. Also, pointers to other ways to run Llama, either on premise or in the cloud ... - Llama Recipes QuickStart- Provides an introduction to Meta Llama using Jupyter notebooks and also demonstrates running Llama locally on macOS. ... - Llama.cpp- Uses the portability of C++ to enable inference with Llama models on a variety - (single_source) Llama Guard 4 shares the same tokenizer and vision encoder as Llama 4 Scout and Maverick. - supports https://llama.com/docs/model-cards-and-prompt-formats/llama-guard-4 > Llama Guard 4 (12B) is our latest safeguard model with improved inference for detecting problematic prompts and responses. It is designed to work with the Llama 4 line of models, such as Llama 4 Scout and Llama 4 Maverick. ... Llama Guard 4 is a natively multimodal safeguard model. The model has 12 billion parameters in total and uses an early fusion transformer architecture with dense layers to keep the overall size small. The model can be run on a single GPU. Llama Guard 4 shares the same tokenizer and vision encoder as Llama 4 Scout and Maverick. ... Llama Guard 4 is also compatible with the Llama 3 line of models and can be used as a drop-in replacement for Llama Guard 3 8B and 11B for both text-only and multimodal applications. However, Llama Guard 3 1B still holds value for specific - (single_source) A Llama 3 prompt can optionally contain a single system message or multiple alternating user and assistant messages, but always ends with the last user message followed by the assistant header. - supports https://llama.com/docs/model-cards-and-prompt-formats/other-models > # Other Models ... - Llama 3 - Llama Guard 2 - Code Llama 70B - Llama Guard 1 - Code Llama - Llama 2 ... ### Llama 3 ... You can find details about this model in the model card. ... Special Tokens used with Llama 3 ... a particular message ... The possible roles can be: system, user, assistant. ... <|end_of_text|>: This is equivalent to the EOS token. On generating this token, Llama 3 will cease to generate more tokens. ... A prompt can optionally contain a single system message, or multiple alternating user and assistant messages, but always ends with the last user message followed by the assistant header. ... Llama 3 ... Note: Newlines (0x0A) are part of the prompt format, for clarity in the example, they have been represented as actual new lines. ... begin_of_text ... Llama 3 Instruct - (single_source) Llama Guard 4 (12B) is a safeguard model designed to detect problematic prompts and responses, and is designed to work with the Llama 4 line of models such as Llama 4 Scout and Llama 4 Maverick. - supports https://llama.com/docs/model-cards-and-prompt-formats/llama-guard-4 > Llama Guard 4 (12B) is our latest safeguard model with improved inference for detecting problematic prompts and responses. It is designed to work with the Llama 4 line of models, such as Llama 4 Scout and Llama 4 Maverick. ... Llama Guard 4 is a natively multimodal safeguard model. The model has 12 billion parameters in total and uses an early fusion transformer architecture with dense layers to keep the overall size small. The model can be run on a single GPU. Llama Guard 4 shares the same tokenizer and vision encoder as Llama 4 Scout and Maverick. ... Llama Guard 4 is also compatible with the Llama 3 line of models and can be used as a drop-in replacement for Llama Guard 3 8B and 11B for both text-only and multimodal applications. However, Llama Guard 3 1B still holds value for specific - (single_source) Llama Guard 4 is compatible with the Llama 3 line of models and can be used as a drop-in replacement for Llama Guard 3 8B and 11B for both text-only and multimodal applications. - supports https://llama.com/docs/model-cards-and-prompt-formats/llama-guard-4 > Llama Guard 4 (12B) is our latest safeguard model with improved inference for detecting problematic prompts and responses. It is designed to work with the Llama 4 line of models, such as Llama 4 Scout and Llama 4 Maverick. ... Llama Guard 4 is a natively multimodal safeguard model. The model has 12 billion parameters in total and uses an early fusion transformer architecture with dense layers to keep the overall size small. The model can be run on a single GPU. Llama Guard 4 shares the same tokenizer and vision encoder as Llama 4 Scout and Maverick. ... Llama Guard 4 is also compatible with the Llama 3 line of models and can be used as a drop-in replacement for Llama Guard 3 8B and 11B for both text-only and multimodal applications. However, Llama Guard 3 1B still holds value for specific - (single_source) The Llama model family includes Llama 3, Llama Guard 2, Code Llama 70B, Llama Guard 1, Code Llama, and Llama 2 as listed on the Other Models page. - supports https://llama.com/docs/model-cards-and-prompt-formats/other-models > # Other Models ... - Llama 3 - Llama Guard 2 - Code Llama 70B - Llama Guard 1 - Code Llama - Llama 2 ... ### Llama 3 ... You can find details about this model in the model card. ... Special Tokens used with Llama 3 ... a particular message ... The possible roles can be: system, user, assistant. ... <|end_of_text|>: This is equivalent to the EOS token. On generating this token, Llama 3 will cease to generate more tokens. ... A prompt can optionally contain a single system message, or multiple alternating user and assistant messages, but always ends with the last user message followed by the assistant header. ... Llama 3 ... Note: Newlines (0x0A) are part of the prompt format, for clarity in the example, they have been represented as actual new lines. ... begin_of_text ... Llama 3 Instruct ## Timeline - 2026-08-28T10:00:07.000Z: Llama Guard 4 (12B) is a safeguard model for detecting problematic prompts and responses, designed to work with Llama 4 Scout and Maverick. Llama Guard 4 is a natively multimodal 12B-parameter model using an early fusion transformer architecture with dense layers. Llama Guard 4 can be run on a single GPU. Llama Guard 4 shares the same tokenizer and vision encoder as Llama 4 Scout and Maverick. Llama Guard 4 is compatible with Llama 3 models and can replace Llama Guard 3 8B and 11B for text-only and multimodal applications. Meta Llama models can run beyond Cloud Service Providers on Linux, WSL, macOS, Jupyter notebooks, and mobile devices. Llama.cpp uses C++ portability to enable Llama model inference on a variety of platforms. Llama Recipes QuickStart introduces Meta Llama via Jupyter notebooks and demonstrates local running on macOS. The Llama model family includes Llama 3, Llama Guard 2, Code Llama 70B, Llama Guard 1, Code Llama, and Llama 2. Llama 3 uses special tokens including an end-of-text token equivalent to EOS, which stops generation. In Llama 3, message roles are system, user, and assistant. A Llama 3 prompt optionally contains a system message or alternating user/assistant messages, always ending with a user message and assistant header.