InternVL2 26B handles both text and images together, processing visual and language inputs in a unified way. At 26 billion parameters, it sits in a mid-to-large range that balances capability with deployability for open-weight models. Being MIT-licensed means teams can use and modify it freely without commercial restrictions.