InternVL3.5-241B-A28B
OpenGVLab · 2025-08-26 · 240.7B parameters
InternVL3.5-241B-A28B is OpenGVLab's 240,699,370,368-parameter open-weight multimodal checkpoint, combining an InternViT-6B vision encoder, a two-layer MLP projector, and a Qwen3-235B-A22B mixture-of-experts language model. It completes the family's CPT, SFT, MPO, and online-RL training pipeline, accepts text, images, and sampled video, and offers a prompt-enabled optional Thinking mode. Its released config exposes 40,960 maximum positions after SFT used 32K sequences; this base checkpoint retains dynamic high-resolution tiling, while the Visual Resolution Router belongs to separate InternVL3.5-Flash variants.
Benchmark scores
| Benchmark | Score |
|---|---|
| OmniDocBench v1.5 | 82.7 |