MiniCPM-V-4_5

MiniCPM-V-4.5 is an 8B-parameter multimodal model from OpenBMB that punches well above its weight class. Built on Qwen3-8B and SigLIP2-400M, it scores 77.0 on OpenCompass across eight vision-language benchmarks, surpassing much larger models including GPT-4o-latest, Gemini-2.0 Pro, and Qwen2.5-VL 72B. OpenBMB positions it as the most performant MLLM under 30B parameters. The model excels at high-resolution image analysis through its LLaVA-UHD architecture, processing images up to 1.8 million pixels with any aspect ratio while using roughly 4x fewer visual tokens than most comparable systems. Video understanding relies on a unified 3D-Resampler that compresses video tokens at a 96x rate, enabling efficient processing of up to 10 FPS video with minimal computational overhead. It reaches leading results on OCRBench and OmniDocBench for document parsing and text recognition, making it well suited to document-heavy workflows. Flexible reasoning is built in with switchable fast and deep thinking modes, letting you trade compute for accuracy on a per-query basis. The model supports 30+ languages and is optimized for on-device deployment through integrations with vLLM, SGLang, Ollama, and llama.cpp, with int4, GGUF, and AWQ quantized variants available. A strong open-weight model suited to vision-language reasoning, document intelligence, video analysis, and resource-constrained deployments.