Llama 3.2 Vision 11b vs Llama 3.2 Vision 90b

Llama 3.2 Vision 11B, released by Meta on September 25, 2024, is the first mid-sized model in the Llama family with vision capabilities, supporting both text and image inputs with text-only outputs. It contains around 11 billion parameters (~10.6B) and features a 128,000-token context window, making it suitable for multimodal reasoning over long documents and image-text tasks. The model was trained on ~6 billion image–text pairs and has a knowledge cutoff of December 2023.

The model is available in a base and an instruction-tuned (“Vision-Instruct”) version, optimized for tasks like captioning, visual question answering, and image reasoning. It leverages Group-Query Attention (GQA) for improved inference efficiency and scalability. While text tasks officially support multiple languages (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai), multimodal (image+text) tasks are supported primarily in English. Llama 3.2 Vision 11B is accessible through Hugging Face, Amazon Bedrock, Azure AI Foundry, NVIDIA NIM, and OCI, making it a widely deployable open-weight multimodal foundation model.

Llama 3.2 Vision 90B, released by Meta AI on September 25, 2024, is the largest vision-capable model in the Llama 3.2 family. With about 90 billion parameters (~88.8B) and a 128,000-token context window, it is designed for high-performance multimodal reasoning over images and text, while producing only text outputs. The model was trained on ~6 billion image–text pairs and instruction-tuned (SFT + RLHF), with a knowledge cutoff of December 2023.

It powers tasks like visual question answering, captioning, and image-grounded reasoning, and achieves strong benchmark performance compared to both open and proprietary models. The model officially supports English for multimodal (image+text) tasks, while text-only inputs extend to eight languages (including German, French, Hindi, and Spanish). Due to its large parameter size, it requires substantial compute resources but is accessible via cloud providers like Amazon Bedrock, Oracle Cloud, and Azure AI Foundry. While highly capable, it is limited to text-only outputs and has stricter multilingual support for vision-based inputs.