MiniCPM-V 4.5 - Faceted Intelligent Open Source 8B Parameter Multimodal Modeling

What is MiniCPM-V 4.5

MiniCPM-V 4.5 is an open source 8B parametric multimodal model of Facade Intelligence, built on Qwen3-8B and SigLIP2-400M, with the ability to process images and videos efficiently. In the visual Token MiniCPM-V 4.5 supports multi-language interaction, and can process 1.8 million pixel images with only 640 visual tokens, which greatly reduces the consumption of computational resources. The model is outstanding in high brush video comprehension, which can receive 6 times the number of video frames and reach 96 times the visual compression rate, which is 12-24 times higher than similar models.MiniCPM-V 4.5 supports multi-language interactions, and can handle more than 30 languages, which is suitable for multi-language customer service and translation scenarios. The document processing ability is also very good, it can handle complex charts and tickets, and supports handwriting OCR and multilingual document parsing. The model supports controlled hybrid reasoning with long and short thinking, and the speed and depth of reasoning can be flexibly adjusted according to actual needs.

Features of MiniCPM-V 4.5

• Efficient visual processingThe visual token consumption is reduced by 75% compared with most models: only 640 visual tokens are needed to process 1.8 megapixel images, which can receive 6 times the number of video frames and achieve 96 times the visual compression rate, 12-24 times more than similar models, with the same visual token overhead.
• multilingual interactionThe multilingual capability supports more than 30 languages, which can be applied to multilingual customer service, multilingual translation and other scenarios.
• Strong document processing skillsBased on the LLaVA-UHD architecture, it can handle high-resolution images of up to 1.8 megapixels in any aspect ratio, and performs very well with handwriting OCR and parsing of complex forms/documents.
• controlled inference: Supports controlled mixed reasoning with long and short thinking, and can flexibly adjust the speed and depth of reasoning according to the actual needs.
• Deployment flexibilityIt provides various quantization model formats such as int4, GGUF, AWQ, etc., which can be selected according to the memory of the device, and supports various deployment methods such as llama.cpp, ollama, vLLM and SGLang.

Core Benefits of MiniCPM-V 4.5

• Outstanding comprehension of high brush videoIt is the industry's first multimodal model with "high-brush" video comprehension capability, which can receive 6 times the number of video frames and achieve 96 times the visual compression rate under the same visual token overhead, which is 12-24 times higher than similar models, and achieves the same size SOTA and exceeds the same size SOTA in the two lists of high-brush video comprehension capability, MotionBench and FavorBench. In MotionBench and FavorBench, which are two lists of high brushed video comprehension capabilities, it reaches the same size SOTA and exceeds the Qwen2.5-VL 72B.
• Excellent performance in image comprehension: Outperforms models such as GPT - 4o - latest in benchmarks such as OpenCompass, efficiently handles high-resolution images, supports images up to 1.8 megapixels in any aspect ratio, excels at handwriting OCR, parsing complex forms/documents, and supports 30+ languages.
• Extensive multi-language supportThe multilingual capability supports more than 30 languages, which can be applied to multilingual customer service, multilingual translation and other scenarios to meet the interaction needs in different language environments.
• Controlled reasoning flexibilitySupport long thinking, short thinking controlled mixed reasoning, according to the actual needs of flexible adjustment of the speed and depth of reasoning, taking into account the efficiency and accuracy.
• Various deployment optionsIt provides various quantization model formats such as int4, GGUF, AWQ, etc., which can be selected according to the memory of the device, and supports various deployment methods such as llama.cpp, ollama, vLLM and SGLang, which is convenient to be used in different devices and scenarios.

What is MiniCPM-V 4.5's official website?

• GitHub repository:: https://github.com/OpenBMB/MiniCPM-V
• HuggingFace Model Library:: https://huggingface.co/openbmb/MiniCPM-V-4_5
• Online Experience Demo:: http://101.126.42.235:30910/

People for whom MiniCPM-V 4.5 is intended

• developers: The model is open source and provides a variety of deployment methods , developers can be based on its secondary development , rapid construction of multimodal applications , such as intelligent customer service , document processing tools .
• research worker: As an open-source model, it is available for researchers to study, analyze and improve, to promote the development of multimodal technology and to explore new application scenarios and algorithm optimization.
• business user: Enterprises can utilize efficient image and video processing capabilities for business scenarios such as surveillance video analysis, product display, and customer service to enhance work efficiency and user experience.
• Mobile device users: The model supports rapid deployment on mobile devices such as iPhone16Pro Max, which is suitable for users who have the need for mobile applications, such as real-time image recognition, document processing, etc., on mobile devices.
• multilingual userSupport for more than 30 languages, suitable for users in multilingual environments, such as multinational corporations, international organizations, etc., to meet the interaction needs of users in different languages.