mistralai/Timestral-209M · Hugging Face

Source: https://huggingface.co/mistralai/Timestral-209M Timestral is afamily of time series foundation modelsfor zero-shot forecasting, available in5M,28M,123M, and209Mparameter versions. This repository contains the 209M checkpoint.

Timestral supports univariate and multivariate forecasting, past-only and known-future covariates, and probabilistic predictions across 21 quantiles. It accepts up to 32,768 historical time steps and forecasts up to 1,024 steps in a single forward pass.

We’re releasing the source code in ourGitHub repositoryto make it easy to run Timestral and reproduce our benchmark results.

https://huggingface.co/mistralai/Timestral-209M#model-detailsModel Details

• **Size:**209M parameters.
• **Architecture:**patch-based transformer with time attention and variate attention, 24 layers, embedding dimension 768.
• **Patching:**32 time steps per patch. Each patch is embedded from its values together with their missing-value indicators.
• **Context length:**up to 32,768 time steps (1,024 patches).
• Forecast horizon: Up to 1,024 time steps (32 patches) in a single forward pass. Longer horizons are forecast block by block, with each block’s median forecast appended to the context.
• **Outputs:**21 quantiles (0.01, 0.05, 0.1, 0.15, …, 0.85, 0.9, 0.95, 0.99).
• **Inputs:**univariate or multivariate target variates, plus past-only and known-future covariates, numeric or categorical (categorical covariates are target-encoded). Missing values are marked as NaN.
• **Inference:**PyTorch on CPU or GPU.

https://huggingface.co/mistralai/Timestral-209M#benchmarksBenchmarks

Timestral-209M achievesstate-of-the-art zero-shot accuracyacross three major forecasting benchmarks:GIFT-Eval,TIME, andfev-bench.

https://huggingface.co/mistralai/Timestral-209M#timestral-209m-compared-with-other-modelsTimestral-209M compared with other models

BenchmarkMetricTimestral-209Mt0-gammaTimesFM-3Toto-2.0-2.5BChronos-2GIFT-EvalMASE ↓0.6580.6610.6670.6960.698CRPS ↓0.4460.4520.4560.4760.485TIMEMASE ↓0.6260.6370.6400.6420.662CRPS ↓0.5220.5340.5360.5390.556fev-benchSQL skill score ↑50.3949.4248.9944.4247.28MASE skill score ↑39.4937.9937.8232.5435.50

https://huggingface.co/mistralai/Timestral-209M#results-across-the-timestral-familyResults across the Timestral family

BenchmarkMetricTimestral-5MTimestral-28MTimestral-123MTimestral-209MGIFT-EvalMASE ↓0.7050.6780.6590.658CRPS ↓0.4790.4610.4480.446TIMEMASE ↓0.6520.6380.6280.626CRPS ↓0.5450.5340.5250.522fev-benchSQL skill score ↑46.1948.4250.1850.39MASE skill score ↑34.5437.2739.1939.49 The Pareto-front plot below illustrates the trade-off between model size and forecasting performance across the Timestral family.

https://huggingface.co/mistralai/Timestral-209M#usageUsage

https://huggingface.co/mistralai/Timestral-209M#installInstall

Set up your environment:

pip install git+https://github.com/mistralai/timestral.git   # Python >= 3.10, torch >= 2.4

Or from a clone withuv, using the versions pinned inuv\.lock:

git clone https://github.com/mistralai/timestral.git && cd timestral
uv sync --frozen
source .venv/bin/activate

https://huggingface.co/mistralai/Timestral-209M#run-the-modelRun the model

Here is a simple example using synthetic data.

import numpy as np
import torch
from timestral import TimestralForecaster

model = TimestralForecaster.from_pretrained(
    "mistralai/Timestral-209M",
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# Synthetic example: one series with 1,000 historical observations.
context = np.random.default_rng(42).standard_normal(1000).cumsum()
forecast = model.forecast(context, prediction_length=96)

median = forecast.median
lower, upper = forecast.quantile(0.1), forecast.quantile(0.9)

Check out ourGitHub repositoryfor more details!

https://huggingface.co/mistralai/Timestral-209M#additional-resourcesAdditional Resources

• Blog post
• GitHub repository
• Tech report: coming soon!

https://huggingface.co/mistralai/Timestral-209M#licenseLicense

Models have been released under a Modified MIT license: Seehere.

You must not use this model in a manner that infringes, misappropriates, or otherwise violates any third party’s rights, including intellectual property rights