Llama-3.2-11B-Vision

Llama 3.2-Vision 11B is a multimodal large language model from Meta's Llama 3.2 family, designed for image reasoning with text output. Built on the Llama 3.1 text backbone with a cross-attention vision adapter, it has 10.6 billion parameters, a 128K token context window, and was pretrained on 6 billion image-text pairs with knowledge cutoff December 2023. The model accepts text and image inputs and generates text responses.

The model targets visual recognition, image reasoning, captioning, and assistant-style chat with images. Instruction-tuned variants excel at visual question answering, document understanding (DocVQA), chart and diagram reasoning (ChartQA, AI2 Diagram), and college-level multimodal problems (MMMU). On text benchmarks, the instruct variant scores 73.0 on MMLU and 51.9 on MATH with chain-of-thought prompting.

Llama 3.2-Vision is released under the Llama 3.2 Community License with gated access on Hugging Face. For image+text applications, English is the officially supported language; text-only tasks support eight languages including German, French, Spanish, and Hindi. The model requires approximately 22GB in BF16, fitting on a single 24GB GPU for inference.