The rapid development of generative artificial intelligence has made it easy to create highly realistic images, which poses a significant risk for the spread of misinformation. However, existing deepfake detection methods are typically confined to identifying specific types of tampered images, with their performance heavily contingent on the design and composition of the training dataset. To address the issue, we propose Instance-Guided Deepfake Detection via Multimodal Large Language Models (MLLM-IGDD), a training-free method that leverages instances to guide multimodal large language models in learning the intrinsic classification logic of image authenticity. Specifically, for each test image, the proposed Instance Recommendation Module (IRM) retrieves a typical authentic image and a deepfake one as in-context references. Furthermore, by integrating domain-informed forensic priors, we introduce the Knowledge-Enhanced Analysis Module (KEAM), which enables the MLLM to detect manipulation-induced anomalies across both visual and semantic dimensions. Experimental results demonstrate that MLLM-IGDD not only accurately assesses image authenticity but also identifies the generation method, localizes manipulated regions, and provides interpretable rationales. It outperforms state-of-the-art methods across diverse datasets while maintaining compatibility with multiple mainstream MLLMs.