Guide · Open source
How to use InternVL: clone, run and self-host an open multimodal model
Updated · written with AI from the project's own documentation and checked against our catalogue data
InternVL is an open-source family of multimodal large language models that take images, and in some cases video, together with text and answer questions or hold a conversation about them. This guide explains what the project includes, what you need before you start, how to clone the repository, and where to look next. By the end you will know whether InternVL fits your work and how to begin.
InternVLOpenGVLab/InternVL · ★ 10.2k · MIT Install options & downloads Quick start
- 1Confirm your hardware. Check that you have a CUDA-capable GPU with enough memory for the model size you want. Larger models need substantial compute.
- 2Install Python and PyTorch. Set up Python and a CUDA-enabled PyTorch build. Look up the exact versions in the project documentation, since the README does not state them.
- 3Clone the repository. Run git clone https://github.com/OpenGVLab/InternVL.git to download the source code.
- 4Read the quick start. Follow the Quick Start section of the README and the documentation at internvl.readthedocs.io to set up the environment.
- 5Load a model and test it. Load one of the published model weights and send it an image with a text question to check the output.
- 6Explore training and evaluation. If you need more, review the training scripts and MPO documentation referenced in the README.
What InternVL is and what it includes
InternVL is positioned in its README as an open-source alternative to commercial models such as GPT-4o. The project publishes model weights, training code and datasets, so you can inspect and reproduce its work rather than only calling a hosted service.
The README references several model generations, including InternVL3.5, InternVL3, and InternVL2.5, along with preference optimization work called MPO (Mixed Preference Optimization). The README is truncated, so specific model sizes and benchmark results are only partly visible in the source.
- Model weights for multimodal chat and question answering
- Training code and shell scripts referenced in the README
- Training datasets such as MMPR and MMPR-Tiny
- A documentation site and a hosted chat demo
Check the prerequisites before installing
The install method is clone-and-build. You need Python and PyTorch, and running the models requires a CUDA-capable GPU. The exact Python and PyTorch versions are not visible in the truncated README, so check the project documentation before setting up your environment.
Larger models need substantial compute. If you do not have GPU hardware or prior machine learning experience, the hosted chat demo linked in the README is a more practical starting point.
- Python installed
- PyTorch with CUDA support, on a machine with a CUDA-capable GPU
- Enough GPU memory for the model size you choose (the README does not give exact figures)
Clone the InternVL repository
The README gives the following command to download the source code. Run it in the folder where you want the project to live.
git clone https://github.com/OpenGVLab/InternVL.gitFind the setup and quick-start instructions
After cloning, follow the quick-start section of the README and the project documentation for environment setup and loading models. The README links to a Quick Start section, a FAQ page, and a full document site at internvl.readthedocs.io, which is the best place to find current installation details.
The repository is organised around the training and chat code. Folders referenced in the README include internvl_chat, internvl_chat_gpt_oss, and shell scripts for specific training stages such as MPO.
- Read the Quick Start section of the README
- Check the documentation at internvl.readthedocs.io for setup details
- Read the FAQ page for common questions
Use InternVL for research, fine-tuning and study
InternVL suits people who want to study how multimodal models are trained and evaluated. The repository exposes the MPO preference optimization pipeline, data construction scripts, and training scripts, which are useful for experiments and teaching.
If you plan to fine-tune or study preference optimization, start with the MPO documentation and the referenced training scripts. Confirm hardware and software requirements in the docs before running any training job.
- Compare open vision-language models on benchmarks
- Study the MPO training pipeline and its data construction scripts
- Fine-tune or evaluate a model on your own image-text tasks
What people use it for
- Comparing open multimodal models against each other on vision-language benchmarks
- Self-hosting a vision-language model instead of paying per call to a commercial API
- Studying how preference optimization (MPO) is applied to multimodal models
- Teaching students how multimodal LLMs are trained and evaluated
Frequently asked questions
Is InternVL free to use?
InternVL is published under the MIT licence, according to the repository. You can read the licence file in the repo to confirm the terms that apply to your use.
What are the requirements to run InternVL?
You need Python, PyTorch, and a CUDA-capable GPU. The README does not list exact versions or memory needs, so check the documentation at internvl.readthedocs.io. Larger models need substantial compute.
How do I install InternVL?
The install method is clone-and-build. Start with git clone https://github.com/OpenGVLab/InternVL.git, then follow the Quick Start and documentation to set up the environment.
Is there a way to try InternVL without installing it?
The README links to a hosted chat demo at chat.intern-ai.org.cn. The README points to this demo for people who want a finished app rather than a self-hosted model.
How does InternVL compare to LLaVA or Qwen-VL?
LLaVA, Qwen-VL, MiniGPT-4 and CogVLM are comparable open projects. The source does not include a head-to-head comparison, so test candidate models on your own task before choosing.
Does InternVL offer production support?
The README does not describe a support commitment. Teams that need guaranteed production support should plan for that separately.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
InternVL is a good fit for researchers, developers who want to self-host a vision-language model, and students studying multimodal training. It is less suitable for people who want a finished consumer app, teams without GPU hardware, or teams that need guaranteed production support, for which a hosted product or another project may serve better.
Install options & downloads