OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Time Moe

Time Moe is a Python implementation of decoder-only time-series foundation models built with a mixture-of-experts architecture. The models generate forecasts autoregressively, with the repository describing context lengths up to 4,096 and arbitrary prediction horizons.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Time Moe — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Time Moe when you need a Python-based forecasting model and are comfortable working with Transformers, tensor inputs and local project dependencies.

Who it is for

  • Developers and researchers building time-series forecasting experiments.
  • Practitioners evaluating or adapting a time-series foundation model.

Documented uses

  • Generate forecasts from numeric time-series sequences.
  • Evaluate model forecasts against prepared benchmark datasets.
  • Prepare JSONL time-series records for fine-tuning.
Before you choose

Requirements & limits

The public GitHub repository identifies an Apache-2.0 license.

  • The documented setup requires Python 3.10 or later and Transformers 4.40.1.
  • The README's inference example loads the 50M model from Hugging Face and requires trust_remote_code.
  • Flash Attention is optional; the project recommends it for faster training and inference with reduced memory use.
  • The README lists covariate support, dynamic-feature fine-tuning and time-series classification as TODO items.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Generate a forecast from an input sequence

Use the documented Transformers example to generate future values from a time-series input.

Before you start

  • Python 3.10 or later and the repository dependencies.
  • Transformers version 4.40.1.
  • A numeric time-series input represented as a tensor; the example uses PyTorch.
  • A model from the documented Maple728/TimeMoE-50M checkpoint.

Steps

  1. Install the repository dependencies with the documented requirements file.

  2. Load the model with Transformers AutoModelForCausalLM and enable trust_remote_code as shown.

  3. Normalize the input sequence and retain its mean and standard deviation, or provide an already-normalized sequence.

  4. Call model.generate with max_new_tokens set to the desired prediction length; reverse the normalization for the generated values when applicable.

Expected outcome

Generated future values appended to the input sequence, with an example that extracts the forecast and reverses normalization.

Guide limitations

  • The documented maximum sequence length is 4,096; the README recommends context length plus prediction length not exceed this.
  • The example uses CPU and notes CUDA as an alternative; Flash Attention is optional.
Documentation 1 ↗
Source-based practical guide

Prepare a dataset for fine-tuning

Format time-series observations as the JSONL structure specified for fine-tuning.

Before you start

  • Time-series observations to include in each record.
  • A JSONL file in which each line is a dictionary with a sequence field containing a list of observations.

Steps

  1. Create one JSONL record for each time series.

  2. Store that series' observations as a list in the record's sequence field.

Expected outcome

A JSONL-formatted dataset matching the input structure described for fine-tuning.

Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
13 October 2026
Original website title
Time-MoE/Time-MoE · Official homepage title · retained in its source language
How we review sourcesSuggest a correction