Google's TimesFM 2.5: A Free Foundation Model for Time-Series Forecasting

While everyone's obsessing over LLMs and image generation, Google Research quietly dropped something massive for anyone working with sequential data. TimesFM 2.5 is a 200M-parameter foundation model purpose-built for time-series forecasting — and it's completely free.

Financial data. Server metrics. Weather patterns. Sales forecasts. If your data moves through time, this model was designed to predict where it goes next.

What Is TimesFM?

TimesFM (Time Series Foundation Model) is a decoder-only foundation model developed by Google Research specifically for time-series forecasting. Think of it like GPT, but instead of predicting the next word, it predicts the next values in a sequence.

The model was published at ICML 2024 and has been iterated since. Version 2.5 is the latest release — and it's a significant upgrade from 2.0.

What Changed in 2.5

The jump from TimesFM 2.0 to 2.5 isn't incremental. They fundamentally rethought the architecture:

200M parameters — down from 500M. Smaller model, better efficiency, runs on more hardware.
16k context length — up from 2,048. That's an 8x increase in how much history the model can see before making predictions.
Continuous quantile forecasting — an optional 30M parameter head gives you confidence intervals (10th to 90th percentile), not just point predictions.
No frequency indicator needed — the old version required you to tell it the data frequency. 2.5 figures it out.
Covariate support via XReg — handle external variables that affect your forecast.

The parameter reduction is the interesting part. Going from 500M to 200M while improving performance means the model is more efficient per parameter. It runs faster, uses less memory, and fits on weaker hardware.

Why This Matters for Builders

Time-series forecasting has historically been painful. You either:

Use classical stats models (ARIMA, Prophet) that require heavy feature engineering
Train custom deep learning models (LSTMs, Transformers) that need massive datasets
Pay for cloud services (AWS Forecast, Google Vertex AI) that get expensive fast

TimesFM gives you a pre-trained foundation model that works out of the box. No feature engineering. No training from scratch. No cloud bills.

Real Use Cases

Where does a time-series foundation model actually help?

SaaS metrics — forecast MRR, churn, DAU with confidence intervals
Infrastructure — predict server load, disk usage, API latency
E-commerce — demand forecasting, inventory planning
Finance — revenue projections, cash flow forecasting
IoT — sensor data prediction, anomaly detection

The quantile forecasting is especially useful. Instead of "we'll hit $100k MRR in June," you get "there's an 80% chance we'll land between $92k and $115k." That's a fundamentally different kind of planning.

Getting Started

It's a Python package. Install, load the model, pass in your data:

# Install
pip install timesfm[torch]

# Use
import timesfm
import numpy as np

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)],
)

It runs on CPU, GPU, TPU, and Apple Silicon. The 200M parameter model is lightweight enough to run on a decent laptop without a GPU.

There's also an official BigQuery integration if you're already in the Google Cloud ecosystem — TimesFM is available as a built-in model in BigQuery ML.

What's Coming

According to the repo, Google is actively working on:

A Flax version for faster inference
Better covariate support
More documentation and notebooks
Support for additional model architectures

The bottom line: If you're working with any kind of time-series data — metrics, sales, usage, finance — TimesFM 2.5 is worth trying. It's free, it's open source, and it's backed by serious research. The 200M parameter model runs on basically anything, and the quantile forecasting gives you something most tools don't: honest uncertainty estimates.