This page provides an overview of the fully-managed Llama models available on Vertex AI, and covers the following topics: Llama models on Vertex AI are available as fully managed, serverless APIs. To use a Llama model, you send a request directly to the Vertex AI API endpoint. With a managed API, you don't need to provision or manage infrastructure. You can stream responses to improve the user experience. A streamed response uses server-sent events (SSE) to incrementally stream the response. The following table provides a high-level comparison of the available Llama model families. The following Llama models from Meta are available in Vertex AI. To access a model, go to its model card in the Model Garden. Models that are in Preview also have a self-deploy option. If you require a production-ready service, use the self-deploy Llama models. Llama 4 Maverick 17B-128E is the largest and most capable Llama 4 model that offers coding, reasoning, and image capabilities. It features a Mixture-of-Experts (MoE) architecture with 17 billion active parameters out of 400 billion total parameters and 128 experts. Llama 4 Maverick 17B-128E uses alternating dense and MoE layers, where each token activates a shared expert plus one of the 128 routed experts. The model is pretrained on 200 languages and optimized for high-quality chat interactions through a refined post-training pipeline. Llama 4 Maverick 17B-128E is multimodal and is suited for advanced image captioning, analysis, precise image understanding, visual questions and answers, creative text generation, general-purpose AI assistants, and sophisticated chatbots that require top-tier intelligence and image understanding. Considerations Llama 4 Scout 17B-16E delivers state-of-the-art results for its size class that outperforms previous Llama generations and other open and proprietary models on several benchmarks. It features MoE architecture with 17 billion active parameters out of the 109 billion total parameters and 16 experts. Llama 4 Scout 17B-16E is suited for retrieval tasks within long contexts and tasks that demand reasoning over large amounts of information, such as summarizing multiple large documents, analyzing extensive user interaction logs for personalization, and reasoning across large codebases. Considerations Llama 3.3 is a 70-billion parameter, text-only model that is instruction-tuned. For text-only applications, it provides enhanced performance compared to Llama 3.1 70B and Llama 3.2 90B. Go to the Llama 3.3 70B model card During the Preview period, you are charged as you use the model (pay as you go). For pay-as-you-go pricing, see Llama model pricing on the Vertex AI pricing page. With Llama 3.2, you can build and deploy generative AI applications that use capabilities such as image reasoning. Llama 3.2 is also designed to be more accessible for on-device applications. Go to the Llama 3.2 90B model card There are no charges during the Preview period. If you require a production-ready service, use the self-hosted Llama models. Considerations When you use Llama 3.1 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. Llama 3.1 405B is Generally Available. You are charged as you use the model (pay as you go). For pay-as-you-go pricing, see Llama model pricing on the Vertex AI pricing page. The other Llama 3.1 models are in Preview. There are no charges for the Preview models. If you require a production-ready service, use the self-hosted Llama models. Go to the Llama 3.1 model card
Available Llama models
Model Family
Description
Key Features
Primary Use Case
Llama 4 Maverick 17B-128E
The largest and most capable Llama 4 model.
Multimodal (text and image), Mixture-of-Experts (MoE) architecture.
Advanced image captioning, visual Q&A, creative text generation, sophisticated chatbots.
Llama 4 Scout 17B-16E
A highly efficient model that delivers state-of-the-art results for its size.
Mixture-of-Experts (MoE) architecture.
Retrieval tasks in long contexts, summarizing large documents, analyzing logs, reasoning across large codebases.
Llama 3.3
A text-only 70B instruction-tuned model with enhanced performance.
Text-only, high performance for its class.
Text-only applications requiring better performance than previous Llama 3 generations.
Llama 3.2
A model designed for image reasoning and accessibility for on-device applications.
Multimodal (text and image), optimized for on-device use.
Applications that require image reasoning.
Llama 3.1
An auto-regressive language model family aligned with human preferences for helpfulness and safety.
Pay-as-you-go availability for the 405B model, tuned with SFT and RLHF.
General-purpose language tasks.
Llama 4 Maverick 17B-128E
Llama 4 Scout 17B-16E
Llama 3.3
Llama 3.2
llama-3.2-90b-vision-instruct-maas
, keep the following in mind:
Llama 3.1
What's next
Fully-managed Llama models
Except as otherwise noted, the content of this page is licensed under the Creative Commons Attribution 4.0 License, and code samples are licensed under the Apache 2.0 License. For details, see the Google Developers Site Policies. Java is a registered trademark of Oracle and/or its affiliates.
Last updated 2025-08-21 UTC.