Llama-3.2-11B-Vision-Instruct

Llama 3.2-11B-Vision-Instruct is Meta's open-weight multimodal instruction-tuned model that accepts text and image inputs and generates text outputs. Built on the Llama 3.1 text backbone with a separately trained vision adapter using cross-attention layers, it is optimized for visual recognition, image reasoning, captioning, document understanding, and assistant-style chat with images. The model was pretrained on 6 billion image-text pairs and instruction-tuned with public vision datasets plus over 3 million synthetically generated examples.

With 11 billion parameters (10.6B reported), 128K context length, and Grouped-Query Attention (GQA), Llama 3.2-Vision delivers strong multimodal performance in a relatively compact footprint. On Meta's official benchmarks, the 11B instruct model achieves 50.7% on MMMU (CoT), 73.0% on MMLU (CoT), 51.9% on MATH (CoT), and 32.8% on GPQA, while excelling at chart and document tasks such as ChartQA (83.4%) and DocVQA (88.4 ANLS). It supports eight languages for text-only tasks and English for image+text applications.

Released on September 25, 2024 under the Llama 3.2 Community License, the model is intended for commercial and research use including visual question answering, document VQA, image captioning, image-text retrieval, and visual grounding. Developers can deploy it via Hugging Face Transformers (MllamaForConditionalGeneration), Ollama, or Meta's native llama codebase, with quantization options enabling inference on consumer GPUs.