Found first: a primary source the press has not covered yet.
Armin Steinhauser has posted TinyCast, a 146,505-parameter probabilistic time-series forecaster that exports to static INT8 and runs end to end on an embedded device without per-signal fitting. On the GIFT-Eval leaderboard it is the only zero-shot entry below 1.4M parameters that emits a full predictive distribution, with no test-data leakage, and the paper reports it defines the size-accuracy frontier for probabilistic accuracy in that setting.
What the source says
TinyCast uses no attention. A zero-parameter spectral detector identifies dominant periods; the context is then folded on their phase before passing through a dilated convolutional encoder and a block-autoregressive quantile decoder. The total parameter count is 146,505. On GIFT-Eval, the paper reports it is smaller than every zero-shot entry whose parameter count can be established. On Chronos-ZS and fev-bench, every neural model that outscores it carries at least 28 times its parameters. The model is available with code and weights on GitHub under a CC-BY 4.0 licence.
Why it matters
Zero-shot time-series forecasting has been dominated by large transformer-based models. TinyCast makes the case that computing periodic structure from the input signal at inference time, via a zero-parameter spectral detector, can substitute for learning it from training data, with the result that the full model fits on a microcontroller as static INT8 code and still produces calibrated distributional forecasts. The practical implication is zero-shot probabilistic forecasting on embedded hardware, a combination the current leaderboard entries do not reach. The 28x minimum parameter gap to every stronger competitor on Chronos-ZS and fev-bench is the source's measure of the efficiency margin that architecture choice achieves.