TServe

Time series serving for foundation models. Load models once, keep them warm, and forecast from any HTTP client or from Python.

Quick start Install TServe

TServe is a server you run, not a hosted API. It loads named time-series models and exposes predictions through JSON, a type-preserving Python client, and a browser dashboard.

• Run anywhere

  ---

  Start with a Docker image, or install with UV or Pip. CPU and GPU options are available.

• Choose your models

  ---

  Serve Chronos, TTM, TimesFM, Moirai, Toto, TiRex, FlowState, Kronos, Mantis, and more.

• Use your preferred client

  ---

  Send JSON from any language, or preserve pandas, polars, pyarrow, and dict inputs with Python.

• Keep models warm

  ---

  Pay model download and load costs at startup instead of on every prediction.

Next steps

• Quick start

  ---

  Launch a server and send your first forecast.

  Quick start

• Installation

  ---

  Choose Docker, UV, Pip, or a source install.

  Install TServe

• Overview

  ---

  What the server loads, and the shape of a predict request.

  How TServe works

• Model catalog

  ---

  Find a model and its matching Docker tag or dependency extra.

  Browse models

• Server

  ---

  Configure model loading, CLI flags, saved models, and the dashboard.

  Run a server

• Clients

  ---

  Send forecasts over HTTP or with the Python client.

  Send predictions