TinyCast Forecaster

A Gradio demo for TinyCast, an attention-free, 146,505-parameter time-series foundation model that forecasts unseen series zero-shot and returns nine quantile forecasts.

How it works

1. Upload a CSV file with a value column (or paste numeric values directly)
2. Select the frequency/domain of your time series
3. Choose a forecast horizon (1–512 steps)
4. Get a probabilistic forecast plot with quantile confidence bands

The model replaces self-attention with dilated causal convolutions and a zero-parameter normalized-periodogram phase prior, so periodicity is computed from the context instead of learned. Every learned operation is a convolution, a matrix multiplication or a normalization, so the model streams in constant memory and runs on CPU.

Model

• Weights: raws-labs/tinycast
• Code: github.com/raws-labs/tinycast
• Paper: arXiv:2608.15767
• License: Apache-2.0