Llama 3.2 90B

About

Llama 3.2 90B is Meta's open-source multimodal large language model that brought vision understanding capabilities to the Llama family for the first time. With 90 billion parameters and a 128K token context window, it processes both text and image inputs, enabling applications like image captioning, visual question answering, document understanding, and multimodal reasoning — all within an open-weight framework that can be self-hosted, fine-tuned, and customized without API dependencies. Llama 3.2 bridges the gap between text-only Llama models and the multimodal capabilities of proprietary models, though its vision capabilities are less refined than GPT-4o's. The model maintains strong performance on text-based tasks like reasoning, coding, and general knowledge that the Llama family is known for. Available under the Llama 3.2 Community License, it can be deployed on consumer GPUs (with sufficient VRAM), fine-tuned for specialized vision-language applications, and integrated into products. The vibrant open-source ecosystem around Llama provides extensive community resources, tools, and model variations. For developers building multimodal applications who want the flexibility of open-weight models, organizations with privacy requirements that prevent sending images to cloud APIs, and researchers working on vision-language AI, Llama 3.2 90B provides a capable open-source option. The main trade-offs are that its vision capabilities lag behind GPT-4o and Claude 4 Sonnet for complex visual reasoning, the 90B size still requires substantial GPU memory, and it has been largely superseded by Llama 4 on text-based benchmarks. For applications where data privacy and model customization are priorities over absolute vision-language performance, Llama 3.2 90B is a strong choice.

Best For

Pricing

Technical Specs

Related Models