Releases · TimeCopilot/foundationforecast

TimesFM 3.0 foundation model: Added support for google/timesfm-3.0-pytorch via a new _TimesFMV3 backend powered by TimesFM3Evaluator and ModelConfig from timecopilot-timesfm>=0.4.0. The wrapper treats each series independently (univariate panel mode), matching the existing TimesFM 2.5 integration pattern.
```
import pandas as pd
from foundationforecast import TimesFM

df = pd.read_csv(
    "https://timecopilot.s3.amazonaws.com/public/data/air_passengers.csv",
    parse_dates=["ds"],
)

model = TimesFM(
    repo_id="google/timesfm-3.0-pytorch",
    alias="TimesFM-3",
    context_length=2048,
)
fcst_df = model.forecast(df, h=12, freq="MS")
```
License note: TimesFM 3.0 pretrained weights are non-commercial; TimesFM 1.0–2.5 checkpoints remain Apache-2.0.
GIFT-Eval: Registered google--timesfm-3.0-pytorch (reference slug TimesFM-3) in the gift-eval model registry and added it to the CI replication subset on m4_weekly/short.
TimesFM family notebook: Added timesfm-family example notebook comparing TimesFM 1.0, 2.0, 2.5, and 3.0 with supported prediction intervals.

Updated README, model hub, and TimesFM docstrings with TimesFM 3.0 support and non-commercial license citation.

Full Changelog: v0.1.2...v0.1.3

Navigation Menu