Mistral Timestral: Open Forecasting Models That Run on CPU

Mistral AI has released Timestral, its first family of time-series foundation models. On October 15 the Paris lab published open weights for four Timestral models, from 5M to 209M parameters, built to forecast how a number will move next: energy demand, store sales, server load, sensor readings. It is a quieter release than a new chatbot. For anyone who runs a business on forecasts, though, it could matter more.

• Four sizes: 5M, 28M, 123M and 209M parameters, all on the same architecture
• Open weights: a Modified MIT license, with models on Hugging Face and code on GitHub
• Long memory: a context of up to 32,768 time steps, with probabilistic output across 21 quantiles
• CPU-friendly: Mistral reports about 10 ms per forecast for the 5M model and about 140 ms for the 209M model on an 8-thread server CPU

What Is a Time-Series Foundation Model?

Most AI headlines are about language models that predict the next word. A time-series foundation model predicts the next value. It reads a history of measurements, such as hourly electricity use or daily orders, and forecasts what comes next.

The "foundation" part means it is pretrained on a huge range of series. It can then forecast data it has never seen without being retrained, which researchers call zero-shot forecasting. Classic forecasting usually means building and tuning a separate statistical model for each dataset. Timestral aims to replace much of that work with a single download.

Timestral's forecasts are probabilistic. It does not give one line; it gives a range of outcomes across 21 quantiles. That matters in practice: a retailer ordering stock or a grid operator planning capacity cares as much about the plausible worst case as the most likely one.

Covariates and Multivariate Forecasting

Timestral is not limited to a single series at a time, a real step up from many earlier open forecasters. According to Mistral, the models handle:

• Multivariate targets: several related series forecast together
• Past-only covariates: signals you know only historically, such as past weather readings
• Known-future covariates: inputs you know ahead of time, such as holidays or scheduled promotions
• Categorical covariates: labels like store region or product type, supported natively

According to the Hugging Face model card, the models use a patch-based transformer with attention across both time and variables, and they forecast longer horizons block by block.

How Does Timestral Compare to Chronos and TimesFM?

Mistral says the 209M model ranks first on all three of the field's main public forecasting benchmarks: GIFT-Eval, TIME and fev-bench. It also says the 209M model beats the previous state of the art with about two-thirds of the parameters. Mistral has submitted its results to the GIFT-Eval and TIME leaderboards, so independent tracking should follow. Until then, treat the ranking as the company's own claim.

Further vendor-reported comparisons, relayed by independent tracker AI TL;DR, say the 28M model beats Chronos-2, a model about four times its size. The same summary says the code is released under Apache 2.0 alongside the Modified MIT weights. The middle sizes may be the most useful in practice. Mistral says the 123M model scores within 0.2% of the 209M model at three quarters of its latency, and the 28M model runs about five times faster than the largest one.

Why CPU Inference Is the Real Story

The speed figures are arguably the headline. Forecasting jobs often run on ordinary servers, edge gateways and factory PCs, not GPU clusters. A model that returns a full probabilistic forecast in tens of milliseconds on a standard CPU fits straight into existing pipelines. Mistral's example is a one-day-ahead forecast from three weeks of 15-minute data.

Organizations that need more control can use Mistral Forge, which adds fine-tuning on proprietary data and supports on-premises, VPC and air-gapped deployment. Companies with sensitive operational data can keep it entirely inside their own walls.

Timestral also joins a growing open-forecasting shelf that includes IBM's Apache-licensed Granite time-series forecaster. It also extends Mistral's open-weight run after Mistral Large 4. For our ongoing coverage of open models, visit our AI news hub.

What This Means for Data Teams

Small teams no longer need a forecasting specialist to get a strong baseline. They can pull a Timestral checkpoint, point it at a CSV of history, and compare the results with their current model in an afternoon. Larger organizations get a permissively licensed model they can fine-tune and deploy wherever their data lives. Forecasting is one of the most widely used kinds of machine learning, and it just got a capable, open and lightweight option.

Sources: Mistral AI — Timestral announcement — October 15, 2026; Hugging Face — Timestral-5M model card — October 2026; AI TL;DR — Mistral Timestral release summary — October 2026.