Speed
Milliseconds per forecast on one CPU, one thread.
Streaming: the cost of one new observation. Bar length is the fourth root of the time, between a log and a linear scale.
| model | streaming | batch, 1 | batch, 16 | batch, 64 |
|---|---|---|---|---|
| laplace (Rust) k=1 | 12 µs | – | – | – |
| laplace (Rust) k=12 | 150 µs | – | – | – |
| laplace (Python) k=1 | 286 µs | – | – | – |
| AutoETS (refit) | 2.0 ms | – | – | – |
| Timestral-28M | 5.5 ms | 5.4 ms | 1.0 ms | 708 µs |
| ARIMA (order fixed, refit) | 7.3 ms | – | – | – |
| AutoTheta (refit) | 7.3 ms | – | – | – |
| Chronos-Bolt-small | 7.3 ms | 7.3 ms | 1.3 ms | 739 µs |
| laplace++ (torch) k=1 | 12.6 ms | – | – | – |
| GARCH-t (refit) | 29.6 ms | – | – | – |
| Timestral-209M | 30.5 ms | 29.7 ms | 4.0 ms | 3.0 ms |
| Prophet (refit) | 30.7 ms | – | – | – |
| t0-gamma | 40.1 ms | 40.6 ms | 4.5 ms | 3.5 ms |
| AutoARIMA (refit) | 62.3 ms | – | – | – |
| TimesFM-3.0 | 66.9 ms | 67.6 ms | 8.6 ms | 5.0 ms |
| TiRex | 94.1 ms | 95.2 ms | 74.4 ms | 74.8 ms |
| Sundial-base-128m | 108.4 ms | 108.5 ms | 33.0 ms | 28.0 ms |
| TimeH2O-v1 | 215.8 ms | 255.4 ms | 155.3 ms | 128.6 ms |
| TimesFM-2.5-200M | 317.1 ms | 322.5 ms | 20.2 ms | 5.0 ms |
Streaming is the cost of the next data point. Batch is the cost of a fresh series, where a pretrained model amortises one forward pass over many series. Hover a column for its definition.
Measured 2026-10-28 on Darwin arm64; built 2026-11-09.