Google Research Releases TimesFM 2.5, a Time Series Forecasting Model

Google Research has announced the release of the latest version of its foundation model for time series forecasting, “TimesFM 2.5.” The model is open-sourced on GitHub and can be installed via PyPI. As a decoder-only foundation model presented at ICML 2024, TimesFM is pre-trained on diverse time series data, enabling zero-shot forecasting capabilities.

Revamped Model Architecture

In TimesFM 2.5, significant architectural changes have been implemented. The number of parameters has been reduced by approximately 60%, from 500M to 200M, improving both model efficiency and inference speed. Meanwhile, the context length has been extended from 2,048 to a maximum of 16,384 (16k), allowing the model to capture longer-term time series patterns.

Additionally, a quantile head with optional 30M parameters has been introduced, enabling continuous quantile forecasts for up to 1,000 steps ahead. This allows the model to evaluate predictive uncertainty as probability distributions rather than mere point estimates, offering high practical value for use cases where confidence intervals are critical, such as financial risk analysis and inventory management.

The previously included frequency indicator has been replaced with a new forecast flag, shifting the model’s design to automatically learn the periodicity of input data.

Installation and Usage

TimesFM 2.5 can be easily installed via PyPI. It supports both PyTorch and Flax (JAX-based) backends, giving users the flexibility to choose based on their environment.

To install the PyTorch version, use the following command:

pip install timesfm[torch] 

For the Flax version:

pip install timesfm[flax] 

To utilize covariate support, add the XReg module:

pip install timesfm[xreg] 

For local development installations, uv-based procedures are available in the official documentation.

Here’s an example of loading the model and performing predictions:

import torch 
import numpy as np 
import timesfm 

torch.set_float32_matmul_precision("high") 

model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(...) 

This code sets PyTorch to high-precision mode and loads the 200M parameter version of TimesFM 2.5. The model weights are retrieved from Hugging Face’s model hub.

Agent Functionality and Fine-Tuning

One of the key advancements in TimesFM 2.5 is its support for agent functionality. Thanks to community contributor borealBytes, the “SKILL.md” guide has been published to demonstrate how AI agents can interface with TimesFM for time series forecasting.

With this feature, it becomes possible to build autonomous time series analysis pipelines integrated with large language models (LLMs). For instance, an LLM can interpret a user’s natural language query, call TimesFM to perform future predictions, and return results along with explanatory text.

In April 2026, an update introduced sample code for fine-tuning using Hugging Face Transformers and PEFT (LoRA). This allows efficient customization of the model for specific domains such as energy demand forecasting, traffic prediction, and financial market analysis. The sample code is located in the timesfm-forecasting/examples/finetuning/ directory.

Additionally, unit tests (tests/) have been incorporated to ensure the quality of core layers, configurations, and utilities. Several community pull requests have also been merged, enhancing the model’s maturity as an open-source project.

Optimizing Inference Performance

The Flax version of TimesFM 2.5 is now available, leveraging JAX’s neural network library. Flax offers optimizations through the XLA compiler and supports TPU environments, making it ideal for large-batch processing and inference with long contexts. Compared to the PyTorch version, noticeable speed improvements can be expected in such scenarios.

Covariate support (XReg) was reintroduced in the October 2025 update. This feature allows the model to incorporate external factors such as temperature, holidays, and economic indicators alongside the target time series. XReg was originally supported in TimesFM 2.0 but was reinstated in version 2.5 following strong community demand.

Integration with Google Products

In addition to its open-source version, TimesFM is also integrated into various Google products. BigQuery ML enables enterprise-level SQL-based time series forecasting, Google Sheets allows direct forecasting within spreadsheets, and Vertex Model Garden provides containerized endpoints for agent calls.

These integrations ensure that users ranging from data scientists to business analysts can easily leverage TimesFM’s forecasting capabilities. Particularly, the integration with BigQuery ML facilitates scalable forecasting for large datasets.

Comparison with Competing Models

In the field of time series forecasting, other open-source tools such as Meta’s Prophet, Amazon’s GluonTS, and Nixtla’s StatsForecast and NeuralForecast are widely used. TimesFM stands out as a foundation model pre-trained on diverse time series data.

While traditional models are domain-specific with limited zero-shot performance, TimesFM’s ability to generalize across finance, energy, weather, demand forecasting, and more without additional training is a unique advantage. Fine-tuning is also an option to further enhance accuracy for specific domains.

The extended context length of 16k differentiates TimesFM from competitors, many of whom handle context lengths in the range of hundreds to a few thousand. Furthermore, the inclusion of quantile forecasting provides practical superiority, as it is not a standard feature in most competing models.

Version History and Community Growth

Since its release in 2024, TimesFM has seen continuous development. After the September 2025 release of TimesFM 2.5, updates such as the addition of the Flax version, reinstated XReg support, agent functionality, fine-tuning samples, and enhanced unit testing have been regularly implemented.

Community-driven contributions have been particularly notable. The AGENTS feature and SKILL.md were developed by community contributors, and the revival of XReg support stemmed from user requests. While Google Research has clarified that this is not an officially supported product, active maintenance continues.

The PyPI package was updated to version 2.0.0 on June 5, 2026, simplifying installation procedures. Older versions (1.0, 2.0) have been archived in the v1 subdirectory and can still be installed using pip install timesfm==1.3.0.

Democratizing Time Series Forecasting

The open-source release of TimesFM is significant for democratizing advanced time series forecasting technology. Historically, building high-accuracy forecasting models required specialized domain knowledge, machine learning expertise, and large amounts of training data. As a foundation model, TimesFM significantly lowers these barriers.

Small businesses and research organizations can adopt it quickly and start forecasting with their own data. The integration with Google Sheets, in particular, allows non-technical users to experience time series forecasting. On the other hand, advanced users can leverage LoRA fine-tuning and agent integration for deeper system integration.

Editorial Opinion

In the short term, the release of TimesFM 2.5 represents a major step forward in the practical application of foundation models for time series forecasting. The balance between reduced parameters and extended context length is particularly noteworthy, achieving both cost-efficient inference and long-term forecasting capabilities. Its integration with enterprise tools like BigQuery ML and Vertex AI could accelerate the adoption of predictive analytics in existing cloud data pipelines. Competitors are likely to ramp up their development of similar foundation models.

From a long-term perspective, Google may position TimesFM as a standard foundation for time series forecasting, integrating it into more products. In a future where autonomous agents handle data analysis, lightweight and versatile models like TimesFM will play a crucial role. However, the fact that TimesFM is not an official Google product may influence enterprise adoption decisions. The sustainability of community-driven development will also be a key factor in its long-term success.

References

• Google Research TimesFM GitHub Repository — Published on 2026-06-19
• TimesFM Hugging Face Collection — Continuously updated
• Google Research Blog: TimesFM — Continuously updated