Llama-3.2-90B-Vision-Instruct

Llama-3.2-90B-Vision-Instruct is Meta's largest multimodal model in the Llama 3.2 family, released in September 2024. Built on the Llama 3.1 70B text backbone with a separately trained vision adapter using cross-attention layers, it accepts text and image inputs and generates text outputs. The model supports a 128K token context window and is optimized for visual recognition, image reasoning, document understanding, chart and diagram analysis, image captioning, and visual question answering.

The model was pretrained on 6 billion image-text pairs and instruction-tuned with publicly available vision datasets plus over 3 million synthetically generated examples. It achieves strong results on multimodal benchmarks including MMMU (60.3%), DocVQA (90.1%), and ChartQA (85.5%), while maintaining competitive text performance with 86.0% on MMLU and 68.0% on MATH. Training used supervised fine-tuning and RLHF for alignment with human preferences for helpfulness and safety.

Llama-3.2-90B-Vision-Instruct is released under the Llama 3.2 Community License for commercial and research use. It is available via Hugging Face Transformers (MllamaForConditionalGeneration), Ollama, AWS Bedrock, and other inference platforms. Note that multimodal use of this model is restricted for individuals and companies domiciled in the European Union.