InternLM-XComposer

Why we included this project

InternLM-XComposer is a practical base for teams that want a vision-language model without building one from scratch. The 7B checkpoint holds long 24K-to-96K interleaved image-text contexts, reads ultra-high-resolution images and fine-grained video, and carries on multi-turn, multi-image conversations; the repository ships the pretrained weights alongside finetuning code and evaluation scripts. Around the core model sit companions worth knowing about, including a multimodal reward model, an OmniLive system built for sustained streaming video and audio interaction, and datasets such as ShareGPT4V and MMDU. That combination makes it a sensible starting point to adapt for visual question answering, document and webpage understanding, or content generation. Treat it as an open foundation model with training recipes rather than a turnkey product.