MiniCPM-V 4.5 was released on 26 August 2025 with 8B total parameters, using a 3D-Resampler and a hybrid reinforcement learning method.
Highlights
77.0 OpenCompass average score — Vision-language benchmark results：Scores an average of 77.0 on OpenCompass, a composite of 8 mainstream benchmarks, at 8B parameters. Comparison models in the same table include GPT-4o-latest, Gemini-2.0 Pro, and Qwen2.5-VL 72B; per-benchmark scores are given there.
Efficient high frame rate and long video understanding：The new unified image-video 3D-Resampler achieves up to 96× video token compression (6 frames of 448×448 only require 64 tokens, while most MLLM requires 1,536), can perceive more video frames without increasing LLM inference costs. It supports high frame rates up to 10 FPS and long-video understanding on Video-MME, LVBench, MLVU, MotionBench, and FavorBench; per-benchmark scores are given in the evaluation table.
OCR, document parsing, and controllable reasoning：Built on LLaVA-UHD, it handles images of any aspect ratio up to 1.8 megapixels (e.g. 1344x1344) using about one quarter the visual tokens of most MLLMs. It scores 89 on OCRBench, against 82.2 for GPT-4o and 88.2 for Qwen2.5-VL-72B in the same table; on OmniDocBench PDF parsing (EN/ZH, lower is better) it scores 0.175/0.253 against 0.233/0.399 for GPT-4o. It supports switchable "fast/deep thinking" modes and 30+ languages; built on RLAIF-V / VisCPM, its MMHal Score is 5.0 and MMHal Rate (lower is better) is 19.4. Per-benchmark scores are given in the evaluation table.
Technical highlights
Unified 3D-Resampler enables high-density video compression：The 3D-Resampler architecture is introduced to jointly compress up to 6 consecutive video frames into 64 tokens, achieving a 96× video token compression rate, supporting high frame rates and long video understanding without increasing LLM computing costs; at the same time. Images, multi-image inputs, and video share a unified encoding, keeping capability and knowledge consistent across the three input forms.
Unified learning paradigm integrates OCR and document knowledge：By dynamically damaging document text areas and requiring model reconstruction, the model adaptively switches between accurate text recognition and multi-modal context reasoning, getting rid of dependence on error-prone document parsers, avoiding hallucinations caused by over-enhancement of OCR data, and achieving top-level OCR. It scores 89 on OCRBench; scores on multimodal knowledge tasks are given in the evaluation table.
Hybrid reinforcement learning optimizes bimodal inference：A hybrid reinforcement learning method jointly optimizes the switchable "fast thinking" and "deep thinking" modes; score changes for both modes are given in the evaluation table. Combining RLPR and RLAIF-V to generalize reasoning from multimodal data, its hallucination metrics are HallusionBench 61.2, MMHal Score 5.0, and MMHal Rate 19.4 (lower is better).
Benchmarks
Integrated multimodality
STEM Reasoning
Document·OCR·Chart
Hallucination review
Multiple pictures·Real world·Instructions to follow
video understanding
The evaluation results are official reports, or local reproductions based on official weights; OmniDocBench is edit distance, and the lower the value, the better.