This study evaluates the capabilities of the Llama 3.2 Vision Instruct model on vision tasks related to Indian traffic data. A custom dataset of 200 images extracted from dashcam footage was created, capturing diverse Indian traffic scenarios. The model was deployed locally and tested using 20 prompts across four categories: vehicle information, human presence, object detection, and road information. Results indicate Llama 3.2 performs consistently in detecting vehicles and humans, with processing times correlating to prompt complexity. The model showed strengths in basic object recognition but faced challenges with more complex scene understanding tasks. Fluctuations in performance highlight areas for potential improvement, particularly in handling the unique complexities of Indian traffic environments. This evaluation provides insights into the model’s applicability for autonomous driving and traffic analysis in the Indian context, while also identifying directions for future enhancements in multimodal vision-language models for specialized domains.