InternVL2.5

InternVL2.5 is an open-source multimodal large language model series from OpenGVLab, released in December 2024. Building on InternVL 2.0's architecture, it improves training strategy, test-time scaling, and data quality across model sizes from 1B to 78B parameters.

The 78B variant was the first open-source MLLM to exceed 70% on the MMMU benchmark, rivaling commercial systems like GPT-4o and Claude 3.5 Sonnet on multidisciplinary reasoning, document understanding, multi-image and video comprehension, and hallucination detection.

The series is fully open with checkpoints on Hugging Face, plus an MPO (mixed preference optimization) alignment variant. It became a go-to open vision-language baseline in early 2025.