TimesFM 2.5 (Transformers)

TimesFM (Time Series Foundation Model) is a pretrained decoder-only model for time-series forecasting. This repository contains the Transformers port of the official TimesFM 2.5 PyTorch release.

Resources and Technical Documentation:

• Original model: google/timesfm-2.5-200m-pytorch
• Paper: A decoder-only foundation model for time-series forecasting
• Transformers docs: TimesFM 2.5

Model description

This model is converted from the official TimesFM 2.5 PyTorch checkpoint and integrated into transformers as TimesFm2_5ModelForPrediction.

The converted checkpoint preserves the original architecture and forecasting behavior, including:

• patch-based inputs for time-series contexts
• decoder-only self-attention stack
• point and quantile forecasts

Usage (Transformers)

import torch
from transformers import TimesFm2_5ModelForPrediction

model = TimesFm2_5ModelForPrediction.from_pretrained("google/timesfm-2.5-200m-transformers")
model = model.to(torch.float32).eval()

past_values = [
    torch.linspace(0, 1, 100),
    torch.sin(torch.linspace(0, 20, 67)),
]

with torch.no_grad():
    outputs = model(past_values=past_values, forecast_context_len=1024)

print(outputs.mean_predictions.shape)
print(outputs.full_predictions.shape)

Conversion details

This checkpoint was produced with:

• script: src/transformers/models/timesfm_2p5/convert_timesfm_2p5_original_to_hf.py
• source checkpoint: google/timesfm-2.5-200m-pytorch
• conversion date (UTC): 2026-02-20

Weight conversion parity is verified by comparing converted-model forecasts against the official implementation outputs on deterministic inputs.

Citation

@inproceedings{das2024a,
    title={A decoder-only foundation model for time-series forecasting},
    author={Abhimanyu Das and Weihao Kong and Rajat Sen and Yichen Zhou},
    booktitle={Forty-first International Conference on Machine Learning},
    year={2024},
    url={https://openreview.net/forum?id=jn2iTJas6h}
}