What's New in TimesFM 2.5

Google Research has updated their TimesFM (Time Series Foundation Model) to version 2.5. This is a decoder-only foundation model specifically designed for time-series forecasting, with the paper published at ICML 2024.

Key Technical Changes

Compared to TimesFM 2.0, the 2.5 model includes several significant updates:

• Parameter count reduced from 500M to 200M
• Context length increased from 2048 to 16k
• Added support for continuous quantile forecast up to 1k horizon via an optional 30M quantile head
• Removed the frequency indicator
• Added new forecasting flags
• Added back covariate support through XReg (as of Oct. 29, 2025 update)

Installation and Setup

The repository is actively being updated with plans for a Flax version for faster inference, more documentation, and notebooks. Current installation requires:

git clone https://github.com/google-research/timesfm.git
cd timesfm
# Create virtual environment with uv
uv venv
source .venv/bin/activate
# Install with torch
uv pip install -e .[torch]
# Or with flax
uv pip install -e .[flax]
# Or with XReg support
uv pip install -e .[xreg]

Basic Usage Example

Here's the basic forecasting workflow from the source:

import torch
import numpy as np
import timesfm

torch.set_float32_matmul_precision("high")
model = timesfm.TimesFM_2p5_200M_torch.from_pretrained("google/timesfm-2.5-200m-pytorch")
model.compile(timesfm.ForecastConfig(
    max_context=1024,
    max_horizon=256,
    normalize_inputs=True,
    use_continuous_quantile_head=True,
    force_flip_invariance=True,
    infer_is_positive=True,
    fix_quantile_crossing=True,
))
point_forecast, quantile_forecast = model.forecast(
    horizon=12,
    inputs=[
        np.linspace(0, 1, 100),
        np.sin(np.linspace(0, 20, 67)),
    ],  # Two dummy inputs
)
Output shapes:
point_forecast.shape → (2, 12)
quantile_forecast.shape → (2, 12, 10): mean, then 10th to 90th quantiles.

Model Availability

The model is available through multiple channels:

• GitHub repository: google-research/timesfm
• Hugging Face collection for all checkpoints
• TimesFM in BigQuery as an official Google product (note: this open version is not officially supported)
• Older versions (1.0 and 2.0) archived in the v1 subdirectory

For developers working with time-series data, this represents a significant update in parameter efficiency and context handling compared to previous versions. The addition of continuous quantile forecasting provides more detailed uncertainty estimates, which is valuable for production forecasting systems.