amine40/Timestral-123M-ONNX

Timestral ONNX

ONNX exports of Timestral-5M, Timestral-28M, Timestral-123M and Timestral-209M, Mistral AI's time series forecasting models, for running in the browser with onnxruntime-web. They power the Timestral in the browser demo.

Precision: float16 weights

The .onnx file stores its weights in float16 to halve the download (tensors under 1,024 values stay float32). They are cast back to float32 when the session loads, so the computation itself runs in float32. Forecasts stay within about 3e-4 of the largest forecast value compared with the full-precision model. These are not the float32 exports. To get those, run scripts/export_onnx.py, which writes both versions.

What the file contains

One decoding pass of the official model, including the rescaling of its output, for any number of series, covariate rows and history lengths (multiples of 32 steps).

Inputs:

name	type	shape
x	float32	rows × T, history, T a multiple of 32
obs	bool	rows × T, false where a value is missing
group_ids	int64	rows, rows with the same id belong to one series
role_ids	int64	rows, 0 target, 1 past-only covariate, 2 known-future covariate
known_future	float32	rows × H, future covariate values, H a multiple of 32 up to 1024
known_future_mask	bool	rows × H

Outputs: quantiles float32, rows × H × 21 (levels 0.01 to 0.99), and last_scale float32, rows.

The rest of the forecast (preparing rows, decoding past 1,024 steps, averaging with the mirrored series, clipping) is in timestral.js. Against the official Python package, forecasts match within 3e-4 of the largest forecast value on the test suite in that repository.

License

The weights are released by Mistral AI under the Modified MIT license in LICENSE, which applies to these exports. Attribution notice: 2026 - Mistral AI.