InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

TL;DR

InternVL3 introduces a unified multimodal pre-training approach that jointly learns visual and linguistic capabilities, achieving state-of-the-art results on open-source MLLMs and maintaining strong language skills.

Contribution

It presents a native multimodal pre-training paradigm with advanced techniques, improving scalability, performance, and alignment over traditional post-hoc methods.

Findings

Achieves 72.2 on MMMU benchmark, setting new state-of-the-art for open-source MLLMs.

Demonstrates competitive performance with proprietary models like ChatGPT-4o and Claude 3.5.

Maintains strong pure-language proficiency while excelling in multimodal tasks.

Abstract

We introduce InternVL3, a significant advancement in the InternVL series featuring a native multimodal pre-training paradigm. Rather than adapting a text-only large language model (LLM) into a multimodal large language model (MLLM) that supports visual inputs, InternVL3 jointly acquires multimodal and linguistic capabilities from both diverse multimodal data and pure-text corpora during a single pre-training stage. This unified training paradigm effectively addresses the complexities and alignment challenges commonly encountered in conventional post-hoc training pipelines for MLLMs. To further improve performance and scalability, InternVL3 incorporates variable visual position encoding (V2PE) to support extended multimodal contexts, employs advanced post-training techniques such as supervised fine-tuning (SFT) and mixed preference optimization (MPO), and adopts test-time scaling strategies alongside an optimized training infrastructure. Extensive empirical evaluations demonstrate that InternVL3 delivers superior performance across a wide range of multi-modal tasks. In particular, InternVL3-78B achieves a score of 72.2 on the MMMU benchmark, setting a new state-of-the-art among open-source MLLMs. Its capabilities remain highly competitive with leading proprietary models, including ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro, while also maintaining strong pure-language proficiency. In pursuit of open-science principles, we will publicly release both the training data and model weights to foster further research and development in next-generation MLLMs.