InternVL3
Open-source multimodal large language model series
InternVL3 is a family of open-source multimodal large language models released in April 2025 by OpenGVLab. It uses native multimodal pre-training to jointly learn vision and language, achieving strong performance on reasoning and text tasks.
- Released in April 2025 by OpenGVLab
- Uses native multimodal pre-training paradigm
- InternVL3-78B scores 72.2 on MMMU benchmark
- Supports video, 3D vision, and GUI agents
- Open weights available for all model sizes
What it is
InternVL3 is a series of multimodal large language models developed by OpenGVLab. Released in April 2025, the family includes variants ranging from 1 billion to 78 billion parameters. The models process text, images, and video, and support tool usage and graphical user interface agents. They are designed for tasks such as industrial image analysis, 3D vision perception, and complex reasoning. The largest model, InternVL3-78B, combines a vision transformer with a Qwen2.5-72B language component.
How it came about
The model follows InternVL 2.5, which was introduced in late 2024. Earlier versions include InternVL 1.0 and 1.5. The development team at OpenGVLab aimed to address alignment challenges found in previous pipelines. Instead of adapting a text-only model after training, they built InternVL3 to learn visual and linguistic skills simultaneously. The research paper describing the architecture and training methods was published on arXiv in April 2025.
Inside the model
InternVL3 uses a ViT-MLP-LLM structure. It connects an InternViT vision encoder to a language model via a projector. A key feature is Variable Visual Position Encoding, which helps the model handle long contexts and flexible visual inputs. The training strategy involves native multimodal pre-training, where image-text and video-text data are mixed with text corpora. This is followed by supervised fine-tuning and mixed preference optimization to improve reasoning. The system also employs test-time scaling to select the best responses during evaluation.
What it can do
The series demonstrates strong performance in multimodal reasoning and mathematics. InternVL3-78B scored 72.2 on the MMMU benchmark, leading among open-source models at the time of release. It handles multi-image inputs, video understanding, and optical character recognition. It competes with proprietary systems like ChatGPT-4o and Claude 3.5 Sonnet in various tasks.
How to use it
OpenGVLab releases the model weights publicly under open terms. The family includes sizes of 1B, 2B, 8B, 9B, 14B, 38B, and 78B parameters. Users can access the models through Hugging Face repositories. A chat demo and API-like interfaces are available for testing. The organization states it will also release training data to support further research and development in the open-science community.
Why it matters
InternVL3 advances the field by proving that joint pre-training of vision and language can outperform sequential adaptation methods. Its open release provides a baseline for researchers studying multimodal alignment. The introduction of variable visual position encoding offers a new technique for managing high-resolution and long-context visual data in transformer architectures.
Questions people ask about InternVL3
What is native multimodal pre-training?
It is a training method where the model learns from both text and visual data simultaneously in one stage. This avoids the need for separate alignment steps used in older approaches, allowing better integration of vision and language skills.
How large is the biggest InternVL3 model?
The largest variant is InternVL3-78B, which has 78 billion parameters. It combines a 6B vision encoder with a 72B Qwen2.5 language model to handle complex multimodal tasks.
Can InternVL3 process video inputs?
Yes, the model supports video data alongside static images and text. It was trained on video-text sequences and can perform tasks related to video understanding and temporal reasoning.
Good reading and sources
There is no English Wikipedia article on InternVL3 yet, so the facts on this page come from the sources listed above: huggingface.co/OpenGVLab/InternVL3-78B, arxiv.org/abs/2504.10479. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.