InternVL3.5

InternVL3.5 is the 2025 generation of Shanghai AI Lab's open InternVL multimodal family, spanning 1B to 38B dense models and MoE variants up to 241B parameters. The headline change is Cascade Reinforcement Learning, a two-stage framework that runs offline preference optimization first for stable convergence, then online RL for refined alignment, which the team credits for stronger reasoning on benchmarks like MMMU and MathVista. A Visual Resolution Router dynamically adjusts visual token resolution, and a Decoupled Vision-Language Deployment strategy splits the vision encoder and language model across GPUs. The team reports roughly 16% better reasoning and 4x faster inference than InternVL3. New capabilities include GUI interaction and embodied agency hooks. Everything is public under Apache 2.0: weights on Hugging Face, code on GitHub, and the training report on arXiv. The 241B MoE variant needs serious hardware, so most developers will use the 4B to 38B range; the 78B-class InternVL3 was already noted for matching GPT-4o on MMMU, and 3.5 builds on that.

Team-reported: +16% reasoning vs InternVL3, 4.05x inference speedup; InternVL3-78B reached 72.2 MMMU (third-party noted)