Timestral: Mistral's time-series foundation models

TL;DR Highlight

A 209M time-series model ranks first on three benchmarks and runs on CPU.

Who Should Read

A backend or data developer who must forecast thousands of sensor, metric or demand series and can't afford training a separate model for each.

Core Mechanics

209M model ranks first on 3 benchmarks
- It ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench.
- It uses about two-thirds of the parameters of the previous state of the art.
- The results come from a single checkpoint without ensembling.
Four sizes, one architecture
- The sizes are 5M, 28M, 123M and 209M, sharing architecture and training recipe.
- The 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME.
- The 209M model beats the larger TimesFM-3 and Falcon-2.0.
Inference runs on CPU, no GPU
- On an 8-thread server CPU, forecasting one day ahead from three weeks of 15-minute history takes 10 ms (5M) and 140 ms (209M).
- The 28M model is five times faster than the 209M model.
- The 123M model is within 0.2% of the 209M score at three quarters of its latency.
Covariates and probabilistic output built in
- It models related series jointly and outputs 21 quantiles to show uncertainty.

Community Reactions

Opinion: TSFMs help when you have 10k series
- A commenter said per-series models have too little history to learn from.
- They said one generic model suffers from scale mismatch and feature work, so a TSFM saves effort.
Doubt that covariates should be the core
- A commenter argued that a model seeing only the last N values loses to a simple model that knows time of day, holidays and weather.
- They also noted a 5M model doesn't know what 'bank holiday' means and just sees a 0 to 1 change.
Adoption and experience reports
- One commenter said their macro team is migrating to this model.
- Another said forecasting models that took months to build five years ago now take minutes.
Skepticism on benchmarks and licence
- One commenter said a 0.03 improvement is drawn to look like a 5x difference.
- Another found it close to the leader t0-gamma but disliked the restrictive licence.
- One noted it is a black box that gives no system equations.

How to Apply

Run zero-shot on one of your series
- Clone the GitHub repo today and run a one-day-ahead forecast with the 5M model on CPU.
- Backtest it on the same window as your current ARIMA or existing model.
Switch alert thresholds to quantiles
- Replace single-value alert rules with upper and lower quantiles from the 21 outputs.
- Check against past incidents whether false alarms drop.
Feed schedules and states as covariates
- Input production schedules, maintenance plans, weekday and machine state as known-future or categorical covariates.
- Compare the error against a run without covariates.
Pick model size by latency and accuracy
- Run 5M, 28M, 123M and 209M on the same data and tabulate error versus latency.
- Choose the largest model that fits your latency budget.

Try It

copy & run

# Paste into Claude or ChatGPT (prompt to plan a Timestral evaluation; the source gives no code so none is included)

I want to evaluate the time-series foundation model Timestral (5M/28M/123M/209M, CPU inference, 21 quantile outputs, past, future and categorical covariates) on our data.
Data description: {{data description: number of series, interval, length}}
Covariate candidates: {{covariate list}}
Current baseline: {{ARIMA etc.}}

Please produce:
1. A backtest split that preserves time order
2. An accuracy and latency comparison table template per model size
3. A way to check how often actual values fall inside the predicted quantiles
4. An experiment design comparing runs with and without covariates

Next step: paste your result table and ask which size to use in production.

Terminology

Time-series foundation modelA model pre-trained on huge data, like an LLM, that forecasts unseen series directly. You don't train one per series.

Zero-shotUsing a model on new data with no extra training. It's like handing someone a task on day one with no onboarding.

CovariateA supporting variable that influences the target but isn't forecast itself, such as holidays, weather or production schedules.

Quantile forecastA forecast giving several low, median and high values instead of one number. Like a rain probability, it shows the width of uncertainty.

Pareto frontThe frontier where no option is better on both criteria, like accuracy and size. Here it means no public model beats it at the same size.