TimesFM 2.5

by Google · 200M parameters · Apache 2.0 · HuggingFace

forecasting quantile outputs zero-shot long context

Google's decoder-only time-series foundation model. Predicts up to 128 timesteps ahead in a single forward pass, with built-in quantile forecasts at 9 confidence levels. Handles up to 16,384 timesteps of context and uses flip invariance to improve prediction stability.

Quick start

$routeframe pull timesfm

Variant	Size	Pull command
f16	441 MB	routeframe pull timesfm
f32	882 MB	routeframe pull timesfm --variant f32

Architecture

Architecture: Decoder-only transformer
Parameters: 200M
Output head: 9-level quantile
Context length: up to 16,384 timesteps
Forecast horizon: up to 128 timesteps in one pass
Positional encoding: RoPE
Flip invariance: Enabled

Run a forecast

$routeframe run timesfm --input "45,48,52,49,55,58,62,59,64,67" --horizon 4

See the CLI docs for advanced usage.