Mistral launches Timestral: time-series forecasting without a GPU

On 15 October 2026, Mistral announced Timestral on its blog. It is the company’s first family of foundation models for time series. This time there is no chatbot and no sonnet in iambic pentameter, just curves: sensor readings, observability metrics, financial volumes. Every company is drowning in them, and nobody has ever put one up at a keynote and waited for applause.

What changes compared with ARIMA

To see why this matters, look at how you forecast a time series today. Classical models such as ARIMA or linear regression are tried and tested. But you fit them on one series at a time, and they only learn from that series’ own history. Deep-learning models such as PatchTST pick up non-linear patterns across many related series. Usually, though, you train them from scratch on each dataset, and what they learn there is no use in another domain.

A foundation model takes the large-language-model recipe and applies it to time series. It is pre-trained on a vast, varied corpus, then forecasts zero-shot: it works on data it has never seen, even where the history is thin. Mistral says it combines three techniques: pre-training on synthetic data, covariate-aware attention and long-context training. If you want to tailor the model to your domain, Mistral offers fine-tuning through Mistral Forge.

Four sizes, and the podium

Timestral comes in four sizes: 5, 28, 123 and 209 million parameters. All four share the same architecture and the same training recipe. The biggest tops out at 209 million parameters, and Mistral itself calls that small. You don’t often hear an AI lab say that about its own model. According to Mistral, it ranks first on the field’s three reference public benchmarks: GIFT-Eval, TIME and fev-bench. It reportedly beats the previous state of the art with two-thirds of the parameters. It also does it with a single checkpoint, without combining several models.

Mistral also offers head-to-head comparisons. The 28M version reportedly beats Chronos-2, roughly four times its size, on GIFT-Eval and TIME. The 209M version reportedly edges out TimesFM-3 and Falcon-2.0 with fewer parameters. Mistral says each size sits on the Pareto frontier between accuracy and size: at the same size, no public model forecasts better. One caveat: every one of these figures comes from the announcement itself. The results have been submitted to the GIFT-Eval and TIME leaderboards, but for now Mistral is still marking its own homework.

No GPU required

This may be the most concrete argument of all: Timestral doesn’t need a GPU to produce its forecasts. Take a standard server CPU with 8 threads and three weeks of history sampled every 15 minutes. Forecasting the next day takes 10 ms with the 5M model and 140 ms with the 209M one. The 28M runs five times faster than the 209M. The 123M comes within 0.2% of the big model’s score, in three-quarters of its response time. So you probably won’t have to pitch a GPU budget to IT leadership, then spend a quarter refreshing a shipping tracker waiting for the cards.

Who it’s for

Mistral is clearly going after businesses: sensor telemetry, capacity planning, energy, supply chains, risk management, maintenance. Its flagship example is a factory fitted with thousands of sensors. On a pump, vibration, temperature and pressure move together. So instead of treating each sensor as a lone wolf, Timestral models these series jointly, which is called multivariate forecasting. Its forecasts are also probabilistic: you get a range of possible values rather than a single number stated with false confidence. It also accepts covariates, either past values or values known in advance, such as a production schedule. On the finance side, it can forecast flows, volumes and volatility across thousands of accounts.

For once, an AI model that predicts the future also tells you how wrong it might be. That’s more honesty than most product roadmaps ship with.