SupraLabs has released SupraWeather-Nano, a preview model designed to classify weather phenomena from raw tabular meteorological data. The architecture utilizes a dedicated Feature Tokenizer and Transformer Encoder, where each input feature receives its own learned token that is aggregated by a CLS token before processing through a small transformer stack. This approach eliminates the need for text inputs or system prompts, allowing users to directly input numerical values to receive a classification result. The model accepts nine specific inputs: temperature, humidity, pressure, pressure trend, wind speed, wind direction, altitude, month, and air mass. It was trained entirely on a synthetic dataset generated by rule-based methods containing 120,000 samples. SupraLabs notes that this is an architecture experiment rather than a tool for real-world forecasting, with five out of six internal stress tests passing successfully.