Recent advancements in weather foundation models—pre-trained on vast amounts of structured numerical data—have set new standards in weather forecasting accuracy. However, their lack of language-based reasoning capabilities leaves a critical opportunity untapped for human-in-the-loop analysis systems. In contrast, large language models (LLMs) excel at understanding and generating text, but they struggle with high-dimensional weather inputs like meteorological datasets. In this work, we take a significant step towards bridging this gap by enabling multimodal LLMs to reason over complex weather data. We address two fundamental challenges: (1) the absence of large-scale, multitask, multimodal datasets for weather reasoning, and (2) the lack of methods for embedding multi-channel weather data into LLM-compatible representations. To tackle these, we introduce a scalable data generation pipeline that constructs diverse question-answer pairs across a wide spectrum of weather-related tasks, from basic lookups to advanced forecasting and extreme event detection. We also leverage pretrained weather foundation models to extract low-dimensional embeddings of weather fields, enabling their integration with LLMs. Our experiments reveal that multimodal weather reasoning is a challenging problem that current models only partially address—highlighting the need for more effective weather representations and richer training data to fully unlock the potential of LLMs in meteorological applications.