Spark was built for tabular analytics. Ray Data is flexible but often requires heavy tuning. Neither reliably handles millions of documents, PDFs, images, or videos. We benchmarked Daft against Spark and Ray Data for common multimodal AI workloads: audio transcription, document embedding, image classification, and video object detection. The results speak for themselves. Daft ran 2-7x faster than Ray Data and 4-18x faster than Spark, reliably finishing jobs while Spark and Ray Data struggled to complete. Read along to learn more about the results, observations we made during these tests, why Spark and Ray Data struggled, and how each engine handles fundamental challenges of multimodal pipelines differently: https://www.daft.ai/blog/benchmarks-for-multimodal-ai-workloads