google-research/timesfm

TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.

What it solves

TimesFM addresses the challenge of time-series forecasting by providing a pretrained foundation model capable of zero-shot generalization. It eliminates the need for per-task tuning when forecasting univariate or multivariate series, even when incorporating dynamic covariates.

How it works

It is a decoder-only foundation model developed by Google Research. The latest version (3.0) supports native multivariate forecasting and handles both past-only and past-and-future covariates. It can be deployed via PyTorch or a native MLX backend for Apple silicon, and is also integrated into Google Cloud services like BigQuery ML and Vertex Model Garden.

Who it’s for

It is designed for data scientists and engineers who need high-performance time-series forecasting across diverse real-world tasks without the overhead of training a model from scratch for every new dataset.

Highlights

  • Top-tier Performance: Ranked #1 overall on fev-bench, TIME Benchmark, and GIFT-Eval benchmarks.
  • Versatile Forecasting: Supports both univariate and multivariate time-series forecasting.
  • Covariate Support: Native handling of past-only and past-and-future dynamic covariates.
  • Flexible Deployment: Available via PyTorch, MLX (for Apple silicon), and Google Cloud enterprise tools.
  • Quantile Forecasting: Provides point predictions as well as continuous quantile forecasts (e.g., 9 deciles).

Written about in

Related

  • Project
    microsoft/TimeCraft

    TimeCraft is a Microsoft‑Research diffusion‑based framework for generating synthetic time‑series data. It supports cross‑domain few‑shot adaptation via a prototype dictionary, natural‑language control, and target‑aware generation that optimises synthetic samples for downstream tasks. Extensions add causal constraints, online augmentation for foundation‑model pre‑training, and continuous‑time generation from irregular observations.

  • Project
    huiliyi37/Tianshu-harness

    A TypeScript-based runtime for coding agents that adds a supervisory layer to ensure tasks are completed with verifiable evidence rather than just model-generated claims.

  • Project
    Nixtla/nixtla

    Nixtla provides a Python SDK that lets you call TimeGPT – a large transformer model for zero‑shot time‑series forecasting and anomaly detection – via a simple API key. Install the package, create a `NixtlaClient`, and use `forecast` or `detect_anomalies` on any pandas DataFrame. The model runs as a hosted service (free‑trial API key available) and supports fine‑tuning, exogenous variables, multi‑series forecasts, Snowflake deployment, and client libraries for other languages. The SDK is Apache‑2.0 open source; the underlying model is closed source.

  • Project
    OpenTSLM/OpenTSLM

    OpenTSLM is an open‑source library that adds native time‑series handling to Llama 3.2 and Gemma LLMs, enabling natural‑language prompting and reasoning over medical signals (ECG, EEG, accelerometer, etc.). It provides pre‑trained checkpoints, a simple `OpenTSLM` Python API, demo scripts for inference on several benchmark tasks, and a curriculum‑learning training pipeline covering MCQ, captioning, and chain‑of‑thought reasoning stages. Install via `pip install opentslm`, load models from Hugging‑Face, and fine‑tune or evaluate using the supplied scripts. Licensed under MIT.

  • Project
    tomas-gajarsky/facetorch

    Facetorch is a Python library for face detection and analysis. It packages open-source deep-learning models into portable, exportable formats and provides a configurable toolkit with components like detectors and predictors.