MiniCPM-V 4.5：通过架构、数据与训练配方打造高效多模态大模型
MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe
3D-Resampler 对图像与视频做高度紧凑的视觉编码，配合混合强化学习在短/长推理模式间切换，为视觉侧提供了一条在 8B 规模下同时压低显存与推理耗时的可复用路线；其视频 token 压缩策略对做视频 MLLM 的效率优化尤其有参考价值。
原文依据 a unified 3D-Resampler model architecture for highly compact encoding over images and videos
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论多模态大语言模型（MLLMs）正快速发展，代表着人工智能发展的前沿。然而，其训练与推理效率已成为制约 MLLM 更易获取、更易扩展的核心瓶颈。为应对这些挑战，我们提出 MiniCPM-V 4.5，一个面向高效率与强性能的 8B 参数模型。我们在模型架构、数据策略与训练方法三方面引入核心改进：一种统一的 3D-Resampler 模型架构，可对图像与视频进行高度紧凑的编码；一种统一的学习范式，无需繁重的数据工程即可同时学习文档知识与文字识别；以及一种混合强化学习策略，使模型在短推理与长推理两种模式下均具备较强能力。OpenCompass 评测的综合实验结果表明，MiniCPM-V 4.5 超越了 GPT-4o-latest 等广泛使用的专有模型，以及 Qwen2.5-VL 72B 等规模显著更大的开源模型。值得注意的是，这一强劲性能是在极高效率下实现的。例如，在广泛采用的 VideoMME 基准上，MiniCPM-V 4.5 以仅 46.7% 的 GPU 显存和 8.7% 的推理时间，超越 Qwen2.5-VL 7B 达 +2.8。
英文原文摘要 1211 字符
Multimodal Large Language Models (MLLMs) are undergoing rapid progress and represent the frontier of AI development. However, their training and inference efficiency have emerged as a core bottleneck in making MLLMs more accessible and scalable. To address the challenges, we present MiniCPM-V 4.5, an 8B parameter model designed for high efficiency and strong performance. We introduce three core improvements in model architecture, data strategy and training method: a unified 3D-Resampler model architecture for highly compact encoding over images and videos, a unified learning paradigm for document knowledge and text recognition without heavy data engineering, and a hybrid reinforcement learning strategy for proficiency in both short and long reasoning modes. Comprehensive experimental results in OpenCompass evaluation show that MiniCPM-V 4.5 surpasses widely used proprietary models such as GPT-4o-latest, and significantly larger open-source models such as Qwen2.5-VL 72B. Notably, the strong performance is achieved with remarkable efficiency. For example, on the widely adopted VideoMME benchmark, MiniCPM-V 4.5 surpasses Qwen2.5-VL 7B by +2.8, using just 46.7% GPU memory and 8.7% inference time.