Paper summary: InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

InternVL3.5 adds Cascade RL (offline then online RL), a Visual Resolution Router and decoupled vision/language deployment, reporting up to +16% reasoning and 4.05× faster inference than InternVL3.

Figure from the paper "InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency"

Figure from Wang et al. (2025), licensed CC BY 4.0. Figure 1 : Comparison between InternVL3.5 and leading MLLMs in general capabilities. Hatched bars represent closed-source commercial models. We report average scores on a set of multimodal general, reasoning, text, and …

Problem

The InternVL series (InternVL 2.5 and 3 were summarised here in Dec 2024 and Apr 2025) keeps pushing open multimodal models. This version targets reasoning and inference efficiency.

Method

Cascade RL: a two-stage, coarse-to-fine process. Offline RL first, for stable convergence, then online RL for refined alignment.
Visual Resolution Router (ViR): dynamically adjusts the resolution of visual tokens without hurting performance.
Decoupled Vision-Language Deployment (DvD): put the vision encoder and the language model on different GPUs to balance the load.

Key results

From the abstract:

up to +16.0% overall reasoning performance and 4.05× inference speed-up compared with InternVL3
new abilities such as GUI interaction and embodied agency
the largest model, InternVL3.5-241B-A28B, is state of the art among open MLLMs on general multimodal, reasoning, text and agentic tasks, narrowing the gap to commercial models like GPT-5 (Figure 1)

Limitations

There’s no dedicated limitations section in the extracted text. Comparisons with commercial models are the authors’ own.

Paper summary: InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

Problem

Method

Key results

Limitations

Sources