Mini-Gemini is a vision-language model framework that extends large language models with multi-modal image understanding and generation capabilities. The framework supports a range of dense and Mixture of Experts (MoE) LLMs from 2B to 34B parameters, enabling image comprehension, visual reasoning, and image generation tasks. Built on the LLaVA architecture, it provides model weights, training code, and inference capabilities through HuggingFace integration.