InternVL2-40B

InternVL2-40B

Quick answer: InternVL2-40B is OpenGVLab's 40-billion-parameter open-weight multimodal model, which led the MMT-Bench VAL split at release, ahead of proprietary models like GPT-4o and Gemini Pro 1.5.

At a Glance

Where InternVL2-40B leads

• Top scorer on the MMT-Bench VAL split at release, ahead of GPT-4o and GeminiPro1.5
• Fully open weights, enabling self-hosted deployment and fine-tuning

Where it lags

• Requires substantial GPU infrastructure to self-host at 40B parameters

Best for: teams needing a strong open-weight multitask multimodal model for broad visual recognition, reasoning, and planning tasks.

What InternVL2-40B Is

InternVL2-40B is part of OpenGVLab's InternVL2 series of open-weight multimodal LLMs. It achieved the top score on the MMT-Bench VAL split, a comprehensive multitask multimodal benchmark spanning 32 meta-tasks and 162 subtasks, outperforming leading proprietary models at the time of evaluation.

Specifications