PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, Zhaokai Wang, Zhe Chen, Hongjie Zhang, Ganlin Yang, Haomin Wang, Qi Wei, Jinhui Yin, Wenhao Li, Erfei Cui, Guanzhou Chen, Zichen Ding, Changyao Tian, Zhenyu Wu, Jingjing Xie, Zehao Li, Bowen Yang, Yuchen Duan, Xuehui Wang, Zhi Hou, Haoran Hao, Tianyi Zhang, Songze Li, Xiangyu Zhao, Haodong Duan, Nianchen Deng, Bin Fu, Yinan He, Yi Wang, Conghui He, Botian Shi, Junjun He, Yingtong Xiong, Han Lv, Lijun Wu, Wenqi Shao, Kaipeng Zhang, Huipeng Deng, Biqing Qi, Jiaye Ge, Qipeng Guo, Wenwei Zhang, Songyang Zhang, Maosong Cao, Junyao Lin, Kexian Tang, Jianfei Gao, Haian Huang, Yuzhe Gu, Chengqi Lyu, Huanze Tang, Rui Wang, Haijun Lv, Wanli Ouyang, Limin Wang, Min Dou, Xizhou Zhu, Tong Lu, Dahua Lin, Jifeng Dai, Weijie Su, Bowen Zhou, Kai Chen, Yu Qiao, Wenhai Wang, Gen Luo

arXiv:2508.18265v2cs.CV

TL;DR

Open-source multimodal models still lag commercial systems on complex multimodal, reasoning, text, and agentic tasks, while increasing visual capability raises inference costs. InternVL3.5 addresses these challenges with Cascade RL for reasoning and ViR plus DvD for efficiency, achieving +16.0% overall reasoning improvement and 4.05× inference speed-up over InternVL3.

  • Problem

    Open-source MLLMs face capability gaps against commercial models on complex tasks, while stronger visual capabilities create a computational-cost bottleneck.

  • Method

    InternVL3.5 combines two-stage Cascade RL with ViR for adaptive visual-token resolution and DvD for separate vision-language deployment.

  • Results

    +16.0% overall reasoning improvement and 4.05× inference speed-up are reported compared with InternVL3, alongside leading open-source performance across broad tasks.

  • Takeaways & Limitations

    InternVL3.5 provides an open-source multimodal model family spanning 1B to 241B with stronger reasoning, versatility, and inference efficiency, with models and code publicly released.

  • Takeaways & Limitations

    InternVL3.5 does not achieve notable improvement over InternVL3 in multimodal understanding, partly because understanding is near saturation and text and reasoning optimization slightly impairs it.

Abstract

from arXiv · show

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.

1 Introduction

InternVL3.5 targets the widening capability and efficiency gap between open-source and commercial multimodal models with Cascade RL, ViR, and DvD. Experiments report leading open-source performance, stronger reasoning, and faster inference across diverse tasks.

  • Motivation: Commercial models have created large gaps over open-source MLLMs on complex multimodal, text, reasoning, and agentic tasks, while rising visual capability increases computational costs.The introduction motivates both capability improvement and inference-efficiency solutions.
  • Cascade RL: InternVL3.5 introduces Cascade RL, combining offline reinforcement learning for stable warm-up with online reinforcement learning for distribution refinement and higher reasoning performance.The offline stage supplies high-quality rollouts for the subsequent online stage, and the framework scales from 1B to 241B models.
  • Efficiency: ViR dynamically selects a visual-token resolution to reduce inference cost with negligible performance sacrifice, while DvD places vision encoders and language models on separate GPUs.Together, these methods provide a hardware-friendly implementation for faster inference.
  • Evaluation: InternVL3.5 consistently leads open-source MLLMs overall and remains competitive with closed-source models across multimodal, text, GUI, embodied, and SVG tasks.The largest model narrows the gap to GPT-5 to 3.9% and surpasses Step-3 by +8.4 in text tasks.
  • Evaluation: +16.0% overall reasoning improvement and 4.05× inference speed-up are reported over InternVL3, with MMMU scores of 73.4 for InternVL3.5-8B and 77.7 for InternVL3.5-241B-A28B.The model family spans 1B to 241B with dense and mixture-of-experts variants, and models and code are publicly released.

2 InternVL3.5

InternVL3.5 combines expanded model architectures and data with staged post-training to improve multimodal capability, reasoning, and efficiency. Its Cascade RL, visual token routing, and decoupled deployment target reasoning quality and inference costs.

  • Model Architecture: InternVL3.5 retains the ViT–MLP–LLM paradigm while spanning multiple model scales with dense and mixture-of-experts variants.The language models are initialized from Qwen3 and GPT-OSS, and the vision encoders from InternViT-300M and InternViT-6B.
  • Model Architecture: InternVL3.5-Flash routes image patches to different compression rates, reducing visual tokens by 50% while maintaining nearly 100% of InternVL3.5’s performance.Patches can be compressed to 256 or 64 tokens according to semantic richness.
  • Post-Training: The training pipeline uses native pre-training, supervised fine-tuning, and Cascade RL, while Flash adds consistency training and router training for ViR.These stages target vision-language alignment, downstream-task adaptation, reasoning improvement, and efficient visual-resolution selection.
  • Post-Training: Cascade RL combines offline and online reinforcement learning, using offline training for stable, efficient improvement before online refinement.The paper reports better stability because offline rollout collection and updates are decoupled, while stronger intermediate models support more stable online training.
  • Infrastructure: DvD separates vision and language processing so ViT computations overlap asynchronously with LLM prefilling and decoding.This reduces blocking and resource conflicts while allowing the language server to focus on language-model computation.

3 Experiments

InternVL3.5 is evaluated across general, reasoning, text-centric, and agentic multimodal tasks, where it shows leading open-source performance and approaches GPT-5 on overall results.

  • Evaluation setup: InternVL3.5 is evaluated across 35 benchmarks spanning general, reasoning, text-centric, and agentic multimodal tasks.The benchmarks are organized into four task categories.
  • Overall comparison: InternVL3.5 consistently maintains a leading position among open-source MLLMs in overall score.The comparison includes flagship InternVL3.5 models, frontier open-source systems, and GPT-5.
  • Overall comparison: 74.1 vs. 74.0: InternVL3.5 achieves a similar overall score to GPT-5 on general multimodal tasks.The reported comparison concerns InternVL3.5 and the state-of-the-art closed-source MLLM GPT-5.
  • Reasoning: 77.7 on MMMU and 82.7 on MathVista demonstrate particularly significant gains in complex multimodal reasoning.The passage attributes these improvements largely to Cascade RL and test-time scaling methodologies.
  • Text and agentic tasks: InternVL3.5 outperforms most open-source models on text-related tasks and shows leading performance on several agentic benchmarks.Agentic results include leading SGP-Bench scores of 69.4 and 70.7, plus strong GUI and embodied-task abilities.

3.2 Multimodal Reasoning and Mathematics

InternVL3.5 achieves state-of-the-art multimodal reasoning performance among evaluated open-source models across multidisciplinary, mathematical, and logical benchmarks, with further gains from parallel thinking.

  • Benchmark results: InternVL3.5 achieves state-of-the-art performance across all evaluated reasoning benchmarks among open-source models.Evaluation covers MMMU, five mathematical reasoning benchmarks, and LogicVista.
  • Benchmark results: More than 10 points: InternVL3.5 improves reasoning performance over comparable-scale InternVL3 models across all model sizes.The comparison is made against the previous generation’s counterparts of comparable scale.
  • Benchmark results: 66.9 overall average: InternVL3.5-241B-A28B leads the evaluated open-source models, followed by InternVL3.5-38B at 66.0.The overall score is the average across the reasoning and mathematics benchmarks listed in Table 3.
  • Test-time scaling: Parallel thinking further improves overall reasoning scores by +2.6%, +2.1%, and +1.8% for the 4B, 8B, and 241B-A28B models.These results indicate additional gains from test-time scaling on reasoning-related tasks.

3.3 OCR, Chart, and Document Understanding

InternVL3.5 is evaluated on nine OCR, chart, and document benchmarks and delivers competitive performance, with scores improving consistently as model size increases.

  • Evaluation setup: InternVL3.5 is assessed across nine benchmarks covering OCR, chart, text, document, and related visual-language understanding.The benchmark suite includes AI2D, ChartQA, TextVQA, DocVQA, InfoVQA, OCRBench, SEED-2-Plus, CharXiv, and VCR.
  • Lightweight models: 76.7 overall average: InternVL3.5-2B surpasses similarly sized InternVL3-2B, which scores 74.7.The overall score aggregates the nine benchmarks, with OCRBench normalized when calculating Overall.
  • Lightweight models: 89.4, 70.8, and 68.0: InternVL3.5-2B scores these values on DocVQA, InfoVQA, and SEED-2-Plus, versus 88.3, 66.1, and 64.6 for InternVL3-2B.The comparisons are reported in the same benchmark order: DocVQA, InfoVQA, and SEED-2-Plus.
  • Scaling: 80.0 to 84.6: overall average scores rise across InternVL3.5 models from 4B through 38B as model size increases.The reported sequence covers the 4B, 20B-A4B, 14B, 30B-A3B, and 38B variants.

3.5 Real-World Comprehension

InternVL3.5 performs competitively on real-world comprehension benchmarks, with stronger results at larger scales and strong comprehensive multimodal performance, while understanding gains over InternVL3 are limited.

  • Real-world benchmarks: InternVL3.5 achieves comparable or superior performance to Qwen2.5-VL, MiniCPM-V-4, and Keye-VL on four real-world comprehension benchmarks.The benchmarks are RealWorldQA, MME-RealWorld, WildVision, and R-Bench.
  • Real-world benchmarks: 57.6, 46.8, 49.2, and 57.4: InternVL3.5-1B scores these values on RealWorldQA, MME-RealWorld, WildVision win rate, and R-Bench.These results are reported for the smallest model variant.
  • Comprehensive multimodal understanding: 87.4 on MMBench v1.1, 81.2 on MMVet, and 77.9 on MMStar: InternVL3.5-241B-A28B shows strong comprehensive multimodal performance.Its reported overall score on this comparison is 84.2.
  • Scope of gains: InternVL3.5 does not achieve a notable improvement over InternVL3 in understanding performance.The passage attributes this partly to saturation and partly to text and reasoning optimization slightly impairing multimodal understanding.

3.7 Multimodal Hallucination Evaluation

InternVL3.5 improves multimodal hallucination handling across model scales while retaining strong visual grounding performance. Its largest variant reaches a new state of the art on the RefCOCO series.

  • Multimodal Hallucination Evaluation: +2.6 overall-score improvement at 2B scale and +1.0 at 38B scale improve multimodal hallucination handling over previous InternVL models.The reported gains span HallusionBench, CRPE, and POPE, though the passage notes minor declines on some results.
  • Visual Grounding: 90.1 overall score on the RefCOCO series gives InternVL3.5-14B a +0.8% gain over InternVL3-14B.The visual grounding evaluation covers RefCOCO, RefCOCO+, and RefCOCOg.
  • Visual Grounding: 92.4 overall score establishes new state-of-the-art performance on the RefCOCO series for InternVL3.5-241B-A28B.The result highlights the model’s potential for applications requiring precise multimodal understanding.

3.9 Multimodal Multilingual Understanding

InternVL3.5 shows broad multimodal capability across multilingual, video, GUI grounding, and online agent evaluations. The results include six-language coverage, strong long-video performance, and competitive GUI-agent scores across platforms.

  • Multimodal Multilingual Understanding: Six-language evaluation across MMMB, Multilingual MMBench, and MTVQA shows consistently high multilingual understanding scores.The languages are English, Chinese, Portuguese, Arabic, Turkish, and Russian, with significant improvements over InternVL3.
  • Video Understanding: +2.2% overall performance for InternVL3.5-1B and +4.9% on LongVideoBench demonstrate gains in video understanding over InternVL3-1B.The model also reports competitive or leading scores across VideoMME, MVBench, MMBench-Video, MLVU, and LongVideoBench.
  • GUI Agent Tasks: +2.6% over UI-TARS-72B on ScreenSpot-v2 shows strong GUI grounding by InternVL3.5-241B-A28B.Its ScreenSpot-v2 score is 92.9 versus 95.2 for Seed1.5-VL.
  • GUI Agent Tasks: +8.3% over Qwen2.5-VL-72B on WindowsAgentArena gives InternVL3.5-241B-A28B the best result among evaluated generalist MLLMs.On WebArena-Lite-v2, it achieves 11.7, compared with 1.9 for GPT-4o.

3.12 Embodied Agent Tasks

InternVL3.5 extends multimodal performance to embodied reasoning, SVG understanding, and SVG generation. Results show especially strong gains for small models and state-of-the-art results among open models at larger capacities.

  • Embodied Agent Tasks: 49.3 overall score on VSI-Bench gives InternVL3.5-1B a +19.6% improvement over its predecessor.The passage identifies VSI-Bench as a popular benchmark for spatial reasoning and reports state-of-the-art performance for this model.
  • SVG Understanding: Nearly 70% accuracy on SGP-Bench for InternVL3.5-30B-A3B and InternVL3.5-38B advances open-model state of the art in SVG understanding.InternVL3.5-4B already surpasses Kimi-VL-A3B and the earlier InternVL3-14B, while the 14B variant surpasses Gemma-3-27B.
  • SVG Generation: 14.56 FID for InternVL3.5-38B on Text2SVG is lower than GPT-4o’s 15.18, establishing stronger open-model SVG generation performance on that metric.The models also establish new state-of-the-art performance among open models on Text2SVG and Img2SVG.

3.14 Evaluation on Language Capability

InternVL3.5 is evaluated across broad language, knowledge, reasoning, mathematics, and coding benchmarks. It generally outperforms corresponding language models, including substantial gains at both small and large scales.

  • Language Capability Evaluation: InternVL3.5 is assessed on general knowledge, linguistic understanding, reasoning, mathematics, and coding benchmarks.The evaluation includes MMLU, CMMLU, C-Eval, GAOKAO-Bench, TriviaQA, NaturalQuestions, C3, RACE, WinoGrande, HellaSwag, BigBench Hard, GSM8K-Test, MATH, AIME24, AIME25, and HumanEval.
  • Language Capability Evaluation: +6.7 overall performance for InternVL3.5-1B over Qwen3-0.6B across 15 of 16 text-related benchmarks.The passage reports that InternVL3.5-1B outperforms Qwen3-0.6B on 15 of 16 evaluated benchmarks.
  • Language Capability Evaluation: +2.3 performance over Qwen3-235B-A22B is achieved by InternVL3.5-241B-A28B.The passage describes this improvement as occurring across the corresponding larger-model comparison.

3.15 Ablation Study

The ablations evaluate Cascade RL, ViR, and DvD across reasoning performance, visual understanding, and inference efficiency. Cascade RL improves results across training stages, while ViR preserves performance and DvD accelerates deployment, especially at higher resolutions.

  • Cascade Reinforcement Learning: Cascade RL improves reasoning progressively beyond SFT and MPO baselines across dense and MoE models.The 2B model gains 12.2% over SFT, while the 241B-A28B model gains 6.5%.
  • Visual Resolution Router: 4.05× speedup is achieved when ViR is combined with DvD, exceeding DvD’s standalone acceleration of up to 2.01×.The efficiency gains also apply to the large InternVL3.5-241B-A28B MoE model.
  • Visual Resolution Router: ViR maintains multimodal understanding and reasoning performance compared with InternVL3.5 without ViR.On 8B DocVQA, the Flash model scores 80.2 versus 79.8 for InternVL3.5.
  • Decoupled Vision-Language Deployment: DvD accelerates inference by up to 2.01 times for InternVL3.5-241B-A28B and 1.97 times for InternVL3.5-38B.DvD benefits both pre-filling and next-token prediction stages.
  • Decoupled Vision-Language Deployment: DvD’s speed-up on InternVL3.5-38B rises from 1.19 to 1.97 as input resolution increases from 448 to 1344.The passage attributes this trend to larger visual computation blocking language-model computation.

4 Conclusion

InternVL3.5 combines Cascade RL for stronger reasoning with ViR and DvD for lower inference cost. It improves overall reasoning performance by 16.0% and inference speed by 4.05× over InternVL3, while its largest model leads open-source MLLMs across several task categories.

  • Conclusion: InternVL3.5 combines Cascade RL with ViR and DvD to improve general performance and inference speed across tasks.Cascade RL combines offline and online reinforcement learning, while ViR and DvD reduce inference cost.
  • Conclusion: 16.0% improvement in overall reasoning performance and 4.05× speed-up are achieved compared with InternVL3.These are the paper’s reported headline gains over the predecessor.
  • Conclusion: InternVL3.5-241B-A28B achieves the highest overall score among leading open-source MLLMs across multimodal, reasoning, text, and agency tasks.The conclusion states that this narrows the gap with top-tier commercial models such as GPT-5.
Loading 2508.18265v2…