Timer (84M)

Timer is THUML's 84M-parameter decoder-only Time-Series Transformer for zero-shot point forecasting. The official model card positions this checkpoint as a general-purpose univariate forecaster with context lengths up to 2880, built on the original Timer architecture rather than the later long-context extension.

Architecturally it is a decoder-only causal transformer that generates forecasts autoregressively, pretrained on 260B time points from diverse domains including energy, traffic, finance, weather, and industrial sensors, and released under Apache-2.0. The causal design suits long-horizon univariate generation across heterogeneous domains, but it also carries some characteristic behavior: in the current TSFM.ai audit loop, direct upstream inference reproduces a tendency to flatten simple repeated seasonal toy probes, so this is best understood as a checkpoint trait rather than a hosted-serving quirk.

On TSFM.ai it is strongest once you provide at least one full 96-point patch of history and validate on your own seasonal regime; very short series are a poor fit. It is the smaller, original-architecture entry point to the Timer line — step up to the billion-scale Timer-S1 when you want native probabilistic quantile output, a much longer context window, and substantially higher capacity from the same lineage.