This repository provides an open-source implementation for training LLaVA-NeXT, a large multimodal model that combines a CLIP vision encoder with large language model backbones (Vicuna-7B, LLaMA3-8B). It includes pretraining and supervised fine-tuning stages with all training data and checkpoints released. The codebase builds on the original LLaVA with minimal modifications, making it easy to reproduce and extend the model for research purposes.