Forecast future values for one or several related time series without training
Improve forecasts using known future events like holidays or promotions
Benchmark forecasting performance on standard leaderboards like GIFT-Eval
Run a quick forecast demo in Google Colab with no local setup
| nx-ai/tirex-2 | penthertz/grimoire | tingxiyu/academic-figure-skill | |
|---|---|---|---|
| Stars | 99 | 98 | 100 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | moderate | easy |
| Complexity | 4/5 | 2/5 | 3/5 |
| Audience | researcher | developer | researcher |
Figures from each repo's GitHub metadata at analysis time.
Pip install is only tested on Linux and macOS, Windows users need the Docker setup, and CUDA needs its own Pixi environment.
TiRex-2 is a pretrained AI model for forecasting time series data, meaning any data that changes over time, like sales numbers, sensor readings, or stock prices. It comes from a research paper and is built by NX-AI. The model can predict what happens next for one variable or many variables at once, and it can optionally take into account extra information such as calendar dates, holidays, promotions, or other known future events that might affect the numbers. What makes it notable is that it works zero-shot, meaning a user does not need to train or fine-tune it on their own data before using it. A single version of the model handles both simple, single-variable forecasting and more complex forecasting involving multiple related variables, and it is built on an architecture designed to work efficiently in streaming settings, where new data keeps arriving over time. The model is relatively small, activating around 38 million parameters for single-variable forecasting and about 44 million more when handling multiple variables together. TiRex-2 builds on an earlier, simpler version of the same model that only handled single-variable forecasting. Installing it is done through Python's pip package manager, with optional extra packages for examples and specific integrations, or through a tool called Pixi for a fully reproducible development setup. It currently only works out of the box on Linux and macOS, though a separate Docker setup covers Windows as well. The README includes a getting started notebook that also runs directly in Google Colab without any local setup, along with example code showing how to load the model, generate a forecast for a set of numbers, and plot the result. There is also a paid Pro version of TiRex-2 offered by the company that adds features not in this open version, including real-time streaming updates, performance optimizations for specialized hardware, the ability to fine-tune the model on private data, and support for classification and regression tasks beyond forecasting. The open-source release here is licensed under Apache 2.0, a permissive license that allows free use, modification, and redistribution, including commercially.
A pretrained AI model that forecasts time series data for one or many variables at once without needing any training on your own data.
Mainly Python. The stack also includes Python, PyTorch, Hugging Face.
Free to use, modify, and redistribute, including commercially, under the Apache 2.0 license.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.