InternVL3-8B

InternVL3-8B is OpenGVLab's 8-billion-parameter vision-language model, built differently from most open VLMs: instead of bolting vision onto a text LLM, it jointly learns multimodal and linguistic capabilities in a single pretraining stage ("native multimodal pretraining"). It pairs an InternViT vision encoder with a Qwen2.5-8B language backbone, uses variable visual position encoding (V2PE) for extended multimodal contexts, and adds SFT plus mixed preference optimization post-training. Released under Apache 2.0, so commercial use is fine.

Benchmarks below are provider-reported from the InternVL3 paper (InternVL3-8B): MMMU 62.7, MathVista 71.6, MathVision 29.3, MathVerse 39.8, DynaMath 25.5, WeMath 37.1, LogicVista 44.1. The authors also report the larger InternVL3-78B hitting 72.2 on MMMU (their claimed open-source SOTA) and claim competitiveness with ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro — an author-framed comparison, not an independent result. Context window is 32,768 tokens (verified from the HF config.json llm_config.max_position_embeddings).

Provider-reported (InternVL3 paper, InternVL3-8B):

• MMMU: 62.7 | MathVista: 71.6
• MathVision: 29.3 | MathVerse: 39.8
• DynaMath: 25.5 | WeMath: 37.1 | LogicVista: 44.1
• InternVL3-78B (same paper): MMMU 72.2 — claimed open-source SOTA