OpenGVLab/InternVL

Based on the README, here's a technical summary: Employs a modular architecture combining vision encoders with language models, enhanced by techniques like Variable Visual Position Encoding and Native Multimodal Pre-Training to handle diverse visual inputs at scale. The family spans 1B–241B parameters with specialized optimization methods including Mixed Preference Optimization (MPO) for preference alignment and Multimodal Test-Time Scaling for improved reasoning. Integrates with Hugging Face transformers, supports both GitHub and HF model formats, and provides open-source training pipelines (CascadeRL, data construction) alongside curated datasets like MMPR for reproducibility.

9,879 stars. No commits in the last 6 months.