Timestral

A person can use it to generate probabilistic forecasts from time series data, including data with missing values.

Availability

123M checkpoint in the repository

Pricing model

Not established

Deployment

Inference with PyTorch on CPU or GPU

Privacy

Not established

Licence

Modified MIT license; must not be used to violate third-party rights

• Timestral-5M weights are available from the Mistral AI organization on the Hugging Face Hub. ^[1]

  The weights are on the Hugging Face Hub, in the Mistral AI organization:

• The 5M model takes 17 ms to forecast one day ahead from three weeks of 15-minute history on a standard server CPU with 8 threads. ^[2]

  On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 17 ms with the 5M model and 113 ms with the 209M model.

• The Timestral pipeline can be deployed on-prem, in a VPC or in an air-gapped environment. ^[2]

  The full pipeline can be deployed on our clients’ hardware, including air-gapped environments.

• Timestral-5M has 5M parameters. ^[3]
• The repository contains the Timestral-5M checkpoint. ^[3]

  This repository contains the 5M checkpoint.

• The 209M checkpoint and source code are publicly available.
• The model accepts univariate or multivariate target variates and past-only or known-future covariates. ^{[1] [5]}

  - **Inputs:** univariate or multivariate target variates, plus past-only and known-future covariates, numeric or categorical (categorical covariates are target-encoded).

• The repository contains the 123M checkpoint. ^[5]

  This repository contains the 123M checkpoint.

• The model supports context lengths of up to 32,768 time steps and forecast horizons of up to 1,024 time steps in a single forward pass. ^[5]

  - **Forecast horizon**: Up to 1,024 time steps (32 patches) in a single forward pass.

• Inference uses PyTorch on CPU or GPU. ^{[1] [5]}

  - **Inference:** PyTorch on CPU or GPU.

• The model is released under a Modified MIT license with a restriction against violating third-party rights. ^{[1] [5]}
• Timestral-123M is a time series foundation model for zero-shot forecasting. ^[5]

  Timestral is a **family of time series foundation models** for zero-shot forecasting, available in [**5M**](https://huggingface.co/mistralai/Timestral-5M), [**28M**](https://huggingface.co/mistralai/Timestral-28M), [**123M**](https://huggingface.co/mistralai/Timestral-123M), and [**209M**](https://huggingface.co/mistralai/Timestral-209M) parameter versions.