Hugging Face researchers have unveiled TinyCast, a compact zero-shot time series forecaster that packs predictive power into just 146,505 parameters. The model is attention-free, relying on a zero-parameter spectral detector to compute dominant periods from the context, then folding the data on their phase before a dilated convolutional encoder and a block-autoregressive quantile decoder generate the forecast distribution.
This design philosophy — computing periodicity rather than learning it — allows TinyCast to be exceptionally small. It is the smallest model on the GIFT-Eval board with a public per-configuration result and no declared test-data leakage. Below 1.4 million parameters, it is the only zero-shot entry that emits a predictive distribution instead of mere point forecasts, and it defines the size-accuracy frontier on probabilistic accuracy.
Despite its size, TinyCast competes with much larger models. On Chronos-ZS and fev-bench, every neural model outperforming it carries at least 28 times its parameter count. Its efficiency stems from using only convolutions and matrix multiplications, enabling static INT8 export and full on-device forecasting without per-signal fitting.
In a demonstration of its embedded-device compatibility, TinyCast ran a complete forecast on a Cortex-M7 microcontroller in 4.08 seconds within 731 KB of RAM, at a cost of about 2% of point accuracy. The weights, code, and training recipe are publicly available, inviting further exploration and adoption.