InternVL3.5 - An open-source multimodal large model from Shanghai AI Lab

What is InternVL3.5?

InternVL3.5 (Shusheng·Wanxiang 3.5) is an open-source multimodal large-scale model from the Shanghai Artificial Intelligence Laboratory. The model features comprehensive upgrades in general capabilities, inference capabilities, and deployment efficiency, offering nine sizes ranging from 1 billion to 241 billion parameters, covering different resource requirements. It includes dense models and expert hybrid models (MoE), and is the first open-source multimodal large-scale model to support the GPT-OSS language model platform. InternVL3.5 employs a Cascade Reinforcement Learning (CRL) framework, significantly improving inference capabilities through a two-stage process of "offline warm-up - online fine-tuning." The flagship model, InternVL3.5-241B-A28B, achieved the highest score of 77.7 among open-source models on the multidisciplinary inference benchmark MMMU, surpassing GPT-5. Leveraging Dynamic Visual Resolution Routing (ViR) and a Decoupled Deployment (DvD) framework, the model's response speed is significantly improved under high-resolution input, with the 38B model achieving a 4.05x increase in throughput.

Main functions of InternVL3.5

• Multimodal sensingIt performs well in multimodal perception tasks such as image and video question answering. The 241B-A28B model outperforms existing open-source models with an average score of 74.1 and is close to the commercial model GPT-5 (74.0).
• Multimodal reasoningIt scored 77.7 points in the multidisciplinary reasoning benchmark MMMU, an improvement of more than 5 percentage points over its predecessor, ranking first among open source benchmarks.
• Text capabilitiesThe model achieves an average score of 85.3 on multiple benchmarks such as AIME, GPQA, and IFEval, placing it at the forefront of open-source benchmarks.
• GUI intelligent agentIt enhances the capabilities of the GUI agent, enabling cross-platform automated operations. For example, it outperforms mainstream open-source models with a score of 92.9 in the ScreenSpot GUI location task.
• Embossed Space DeductionIt has stronger grounding capabilities, can be generalized to new and complex embodied scenarios, and supports generalizable long-range object grasping operations.
• Vector graphics processingIt achieved a new open-source score of 70.7 on SGP-Bench, demonstrating its effectiveness in professional scenarios such as web page graphics generation and engineering drawing parsing.

Technical Principles of InternVL3.5

• Cascade Reinforcement LearningThe process involves two stages: offline warm-up and online fine-tuning. In the offline stage, the Mixed Preference Optimization (MPO) algorithm is used to quickly improve basic inference capabilities and provide high-quality samples for subsequent training. In the online stage, the GSPO algorithm is used to dynamically adjust the output distribution based on the samples generated by the model itself, which significantly improves training stability and inference performance.
• Dynamic Visual Resolution Routing (ViR)The compression rate is dynamically selected for each image slice, retaining high resolution in semantically dense areas and adaptively compressing background areas to reduce visual tokens, significantly improving inference speed while almost without sacrificing performance.
• Decoupled Deployment Framework (DvD)By placing the visual encoder and language model on different GPUs, and combining BF16 precision feature transfer with asynchronous pipeline design, visual computing and language generation can be executed in parallel, significantly improving throughput and solving the resource blocking problem of traditional serial deployment.
• Full-scale model optimizationIt offers nine model sizes ranging from 1 billion to 241 billion parameters, covering different resource requirements and including dense models and expert hybrid models (MoE). It is the first open-source multimodal large model to support the GPT-OSS language model base, meeting diverse application needs.
• Multimodal collaborative reasoningBy integrating multi-dimensional information such as vision and language, it enables efficient processing of multimodal tasks, enhances the model's reasoning ability in complex tasks, and promotes the leap of multimodal technology from "understanding" to "action".

Project address for InternVL3.5

• Github repositoryhttps://github.com/OpenGVLab/InternVL
• HuggingFace model addresshttps://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B
• Technical Reporthttps://huggingface.co/papers/2508.18265
• Online experience addressThe Scholar Model

Application scenarios of InternVL3.5

• Office AutomationThrough the GUI intelligent agent function, cross-platform automated office operations can be achieved, such as automatically processing Excel data entry, PPT design and layout, email sending and other tasks, thereby improving work efficiency.
• Smart Home Control: Utilizing embodied spatial reasoning capabilities, it assists robots in completing tasks such as object location, path planning, and physical interaction in the home environment. For example, intelligent cleaning robots autonomously plan cleaning paths based on the layout of the environment.
• Educational guidanceSupported by multimodal reasoning and textual capabilities, it provides students with personalized learning guidance, answers complex subject-specific questions, such as mathematical and physical problem-solving, and logical reasoning training, thereby improving learning outcomes.
• Content creationLeveraging general multimodal perception capabilities, it automatically generates image descriptions, video subtitles, and other content, helping creators quickly generate creative content and improving the efficiency and quality of content creation.
• Web Design and Graphics GenerationUtilizing vector graphics processing capabilities, it generates or edits SVG vector graphics according to user instructions, suitable for web design, icon creation, etc., improving design efficiency and personalized experience.