mrfakename/Timestral-209M-ONNX

Timestral-209M ONNX

ONNX export of the network in mistralai/Timestral-209M, made for running in the browser with onnxruntime-web on WebGPU. Used by the mrfakename/timestral-209m-webgpu Space.

file	weights	size
onnx/model.onnx	fp32	841 MB
onnx/model_fp16.onnx	fp16 (attention scores, softmax and norms in fp32)	422 MB

What the graph covers

One forward pass for a single series without covariates:

• inputs: scaled float32 (1, T), the context already normalized (causal per-patch mean/std, then asinh), T a multiple of 32, up to 32768; observed bool (1, T); horizon float32 (H,), only its length is read (patches to decode, 1 to 32).
• output: quantiles float32 (1, H * 32, 21), the 21 quantiles (0.01 … 0.99) in the normalized space of the last context patch.

Scaling, de-normalization, block-by-block decoding past 1024 steps, the refine frame, symmetric averaging and non-negative clipping are left to the caller, as in timestral/forecaster.py from github.com/mistralai/timestral. A JavaScript port of that wrapper is in the Space (forecast.js); it matches the Python package to about 1e-4 with the fp32 graph.

export.py is the export script (torch.onnx dynamo export of the official package). For fp16 safety it uses an explicit attention with fp32 scores and an exactly rescaled RMSNorm; the math is unchanged.

License

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