Llama 3.2 11B Vision Instruct

Llama 3.2 11B Vision Instruct - Multimodal Image Reasoning Model

Llama 3.2 11B Vision Instruct is Meta's open-weight multimodal model that accepts text and image inputs and produces text outputs. It is architecturally distinct from the text-only Llama 3.2 series: it is built on top of the pretrained Llama 3.1 language model and extends it through a separately trained vision adapter rather than a natively multimodal architecture.

Intended Use Cases

Documented applications include visual recognition, image reasoning, captioning, assistant-like chat with images, Visual Question Answering (VQA), and Document Visual Question Answering (DocVQA) - understanding both text and layout of documents such as maps or contracts.

Safety Design

Meta documented dedicated mitigations specifically for the risk of uniquely identifying individuals in images - a concern specific to multimodal input. The recommended companion safeguard is Llama Guard 3-11B-Vision, which provides vision-aware input and output classification.

Training

Post-training alignment used SFT and RLHF.