RESEARCH & ANALYTICS · WORTH A LOOK
Timestral (Mistral AI)
Mistral's first family of open-weight time-series foundation models (5M, 28M, 123M and 209M parameters) for zero-shot forecasting: give it historical…
What Timestral (Mistral AI) does
Mistral's first family of open-weight time-series foundation models (5M, 28M, 123M and 209M parameters) for zero-shot forecasting: give it historical data such as sales, demand, energy use or sensor readings and it predicts what comes next with 21 probability quantiles, without training a model per dataset. Univariate or multivariate, supports past-only and known-future covariates (e.g. holidays, schedules), up to 32,768 history steps and 1,024 forecast steps per pass. Runs on CPU or GPU via a Python/PyTorch library. Mistral claims #1 among pretrained/zero-shot models on GIFT-Eval, TIME and fev-bench and ~140 ms per forecast for the 209M model on an 8-thread CPU (vendor figures).
Good for
- Forecast sales, demand and inventory needs from past data without building a custom model
- Forecast energy use, capacity or cash-flow-style series with uncertainty ranges for planning
- Spot anomalies in sensor or operations data by comparing actuals against the forecast range
Try this in 10 minutes
- Forecast sales, demand and inventory needs from past data without building a custom model. Use the free plan: pick one real example from your own business and try just that.
- Forecast energy use, capacity or cash-flow-style series with uncertainty ranges for planning. Put the result next to how you do it today. Is it faster, cheaper or better?
- Spot anomalies in sensor or operations data by comparing actuals against the forecast range. Note what you had to fix by hand before you'd show it to a customer.
Mini-tasks based on common small-business uses for this tool.
Honest take
Some claims not independently checked
- Needs Python >= 3.10 and PyTorch; not a no-code app.
- Not deployed by any Hugging Face Inference Provider at time of check.
What we couldn’t verify
- Benchmarks and speed are Mistral's own claims; results submitted to GIFT-Eval and TIME leaderboards.
Pricing
Free to download and self-host: weights under a Modified MIT licence (not usable by companies with over $20M global monthly revenue without a commercial licence from Mistral); inference code Apache 2.0. Fine-tuning/enterprise deployment via Mistral Forge: pricing Unknown.
Prices as listed by the maker when we last checked (2026-10-19); usually in US dollars. Check the website for current prices.
Latest update
- 2026-10-15: Timestral 1.0 released (4 sizes, GitHub inference code, Hugging Face weights); technical report still to come.
Is Timestral (Mistral AI) worth it? Quick answers
Is Timestral (Mistral AI) worth it?
Timestral (Mistral AI) is worth a look if it matches a job you already do (Research & analytics). There is a free tier, so you can test it before paying. Some of the maker's claims have not been independently checked.
How much does Timestral (Mistral AI) cost?
Free to download and self-host: weights under a Modified MIT licence (not usable by companies with over $20M global monthly revenue without a commercial licence from Mistral); inference code Apache 2.0. Fine-tuning/enterprise deployment via Mistral Forge: pricing Unknown.
What is Timestral (Mistral AI) used for?
Mistral's first family of open-weight time-series foundation models (5M, 28M, 123M and 209M parameters) for zero-shot forecasting: give it historical… Typical uses: forecast sales, demand and inventory needs from past data without building a custom model; forecast energy use, capacity or cash-flow-style series with uncertainty ranges for planning; spot anomalies in sensor or operations data by comparing actuals against the forecast range.