Post
67
TinyCast is a 146,505-parameter time series foundation model. Univariate, zero-shot, nine quantiles per step, and small enough to run on a microcontroller.
It is attention-free: ten dilated causal convolution blocks, with periodicity computed from the context by a Fisher test on the periodogram instead of learned. That is most of why it stays small.
We ran it for a full year against a national grid operator's own day-ahead forecast for Belgian electricity demand. TinyCast was the more accurate of the two on 45% of days, reading nothing but 2048 past values. No weather, no calendar, no fitting to that series.
Deployed on an STM32H7 as a static INT8 graph it needs 138 KB of weights, 731 KB peak RAM, and 4.1 s per full forward call.
Thanks to @multimodalart , who built this Space unprompted and handed it over, you can now try it in your browser: raws-labs/tinycast-forecaster
Model: raws-labs/tinycast
Paper: https://arxiv.org/abs/2608.15767
Code, and the grid example: https://github.com/raws-labs/tinycast
It is attention-free: ten dilated causal convolution blocks, with periodicity computed from the context by a Fisher test on the periodogram instead of learned. That is most of why it stays small.
We ran it for a full year against a national grid operator's own day-ahead forecast for Belgian electricity demand. TinyCast was the more accurate of the two on 45% of days, reading nothing but 2048 past values. No weather, no calendar, no fitting to that series.
Deployed on an STM32H7 as a static INT8 graph it needs 138 KB of weights, 731 KB peak RAM, and 4.1 s per full forward call.
Thanks to @multimodalart , who built this Space unprompted and handed it over, you can now try it in your browser: raws-labs/tinycast-forecaster
Model: raws-labs/tinycast
Paper: https://arxiv.org/abs/2608.15767
Code, and the grid example: https://github.com/raws-labs/tinycast