Mistral releases Timestral-209M time-series forecasting model
Original title: mistralai/Timestral-209M — new model trending #29 on Hugging Face
The 209M checkpoint forecasts up to 1,024 steps from 32,768 history steps across 21 quantiles.
Mistral AI published Timestral-209M, the largest checkpoint in a time-series foundation model family that also includes 5M, 28M, and 123M versions. It supports univariate and multivariate targets, past-only and known-future covariates, and probabilistic forecasts. The company reports leading zero-shot scores on GIFT-Eval, TIME, and fev-bench. Weights use a Modified MIT license; a tech report is still forthcoming.
Unconfirmed: State-of-the-art scores are Mistral's own comparisons; an independent tech report is not yet published.
- The model is a 24-layer patch transformer with embedding dimension 768.
- Each patch covers 32 time steps plus missing-value indicators.
- Reported GIFT-Eval scores are MASE 0.658 and CRPS 0.446.
- Inference uses PyTorch on CPU or GPU; Python 3.10 and torch 2.4 are required.
- Source code is at github.com/mistralai/timestral.
- Vendors
- Mistral AI
- Products
- Timestral
- AI models
- Timestral-209M
From the original article
3 min read at huggingface.cotime-series-forecasting · 2,601 downloads · 170 likes · license other
By mistralai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.