We introduce InternVL 3.5, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the Cascade Reinforcement Learning (Cascade RL) framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a Visual Resolution Router (ViR) that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled Vision-Language Deployment (DvD) strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0\% gain in overall reasoning performance and a 4.05$\times$ inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks -- narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.
核心贡献 · Key contributions
提出级联强化学习，结合离线和在线强化学习，实现稳定且可扩展的推理能力提升。 Proposes Cascade RL combining offline and online RL for stable and scalable reasoning enhancement.
提出解耦视觉-语言部署（DvD），分离视觉和语言处理以加速推理。 Presents Decoupled Vision-Language Deployment (DvD) to separate vision and language processing for faster inference.
在通用多模态、推理、文本和智能体任务中，取得开源多模态大语言模型最佳结果。 Achieves state-of-the-art results among open-source MLLMs on general multimodal, reasoning, text, and agentic tasks.
缩小与 GPT-5 等领先商业模型的性能差距，推理性能提升高达 16.0%，速度提升 4.05 倍。 Narrows performance gap with leading commercial models like GPT-5, with up to +16.0% reasoning gain and 4.05x speedup.
支持 GUI 交互、具身智能体和 SVG 理解/生成等新能力。 Supports novel capabilities including GUI interaction, embodied agency, and SVG understanding/generation.
局限 · Limitations
在 HallusionBench 等幻觉基准上的性能仍需改进。 Performance on hallucination benchmarks like HallusionBench still requires improvement.
级联强化学习可能无法在不进一步调优的情况下泛化到所有模型规模或任务类型。 Cascade RL may not generalize to all model scales or task types without further tuning.
ViR 和 DvD 针对特定硬件优化，不同平台上的收益可能有所不同。 ViR and DvD are optimized for specific hardware; benefits may vary on different platforms.
模型的理解性能接近饱和，限制了在某些基准上的进一步提升。 The model's understanding performance approaches saturation, limiting further gains on some benchmarks.
多语言和视频任务的训练数据与策略可能未覆盖所有语言或场景。 Training data and strategies for multilingual and video tasks may not cover all languages or scenarios.
论文章节 · Sections（共 25）
摘要Abstract
1 引言1 Introduction
2 InternVL3.52 InternVL3.5
2.1 模型架构2.1 Model Architecture
2.2 预训练2.2 Pre-Training
2.3 后训练2.3 Post-Training
2.4 测试时扩展2.4 Test-Time Scaling
2.5 基础设施2.5 Infrastructure
3 实验3 Experiments
3.1 与其他先进多模态大语言模型的整体比较3.1 Overall Comparison with Other Advanced MLLMs
3.2 多模态推理与数学3.2 Multimodal Reasoning and Mathematics
3.3 OCR、图表与文档理解3.3 OCR, Chart, and Document Understanding