InternVL3 38B handles both text and images, processing visual and language inputs together in a single model. At 38 billion parameters, it sits in the upper-mid range of open-weight multimodal models, offering substantial capacity for vision-language tasks without requiring proprietary access. It follows the instruct-tuned pattern, meaning it's shaped for conversational and task-following use rather than raw completion.