Which Timestral size should I use?

Mistral says Timestral-123M comes within 0.2% of the 209M model's score at about three-quarters of its latency, and that the 28M model is about five times faster than the 209M model. The 5M model is the fastest at 10 ms per forecast on an 8-thread server CPU, so smaller sizes suit tight latency budgets.

Is Timestral open source?

Mistral AI publishes the Timestral inference code on GitHub under Apache 2.0 and the model weights on Hugging Face under a Modified MIT license. All four checkpoints, from Timestral-5M (21 MB) to Timestral-209M (837 MB), can be downloaded and run on your own servers, including air-gapped setups.

How long a history can Timestral read?

Timestral reads up to 32,768 past time steps and forecasts up to 1,024 steps in one forward pass. Longer horizons are forecast block by block. The models accept past-only covariates and known-future covariates, such as a holiday calendar, alongside the target series, and missing values can be passed as NaN.

How does Timestral compare to TimesFM and Chronos?

Mistral reports that Timestral-209M beats TimesFM-3 and Falcon-2.0 while using fewer parameters, and that Timestral-28M beats Chronos-2 on GIFT-Eval and TIME. On the Hugging Face model card, Timestral-209M scores a MASE of 0.658 and CRPS of 0.446 on GIFT-Eval, where lower is better.

Mistral AI · 2026-10-15 · major

Timestral — Mistral's first open models for time-series forecasting

Timestral is Mistral AI's first family of time-series foundation models, in 5M, 28M, 123M and 209M sizes. The models give zero-shot probabilistic forecasts, and Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench.

Mistral AI announcement artwork for the Timestral time-series forecasting models — Mistral AI

Mistral's tiny Timestral models forecast any time series zero-shot, from 10 ms on a CPU.

Quick facts

Maker	Mistral AI
Sizes	5M, 28M, 123M, 209M parameters
Context	32,768 time steps
Output	21 forecast quantiles, up to 1,024 steps per pass
License	Modified MIT (weights), Apache 2.0 (code)
Availability	Hugging Face, GitHub, Mistral Forge

What is it?

Timestral is a family of four time-series foundation models from Mistral AI, released on 15 October 2026. The models forecast future values of one or many related series, such as sales or sensor readings, without training on your data first. Each forecast is a spread of 21 quantiles, not just one line.

How does it work?

Most of Timestral's training data is synthetic: simple processes with trends, seasonality and noise, combined into multivariate series. A method Mistral calls Decoupled Horizon Forecasting keeps the full history in the sequence and adds learned forecast tokens. Covariate-aware attention gives each series a role, so information flows only from covariates to the targets.

Why does it matter?

Forecasting models are often run on CPUs inside existing data pipelines, so size matters. Mistral reports the 5M Timestral model answers in 10 ms and the 209M model in 140 ms on an 8-thread server CPU, and says the 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME.

Who is it for?

data scientists, forecasting and ops teams

Frequently asked questions

Which Timestral size should I use?: Mistral says Timestral-123M comes within 0.2% of the 209M model's score at about three-quarters of its latency, and that the 28M model is about five times faster than the 209M model. The 5M model is the fastest at 10 ms per forecast on an 8-thread server CPU, so smaller sizes suit tight latency budgets.
Is Timestral open source?: Mistral AI publishes the Timestral inference code on GitHub under Apache 2.0 and the model weights on Hugging Face under a Modified MIT license. All four checkpoints, from Timestral-5M (21 MB) to Timestral-209M (837 MB), can be downloaded and run on your own servers, including air-gapped setups.
How long a history can Timestral read?: Timestral reads up to 32,768 past time steps and forecasts up to 1,024 steps in one forward pass. Longer horizons are forecast block by block. The models accept past-only covariates and known-future covariates, such as a holiday calendar, alongside the target series, and missing values can be passed as NaN.
How does Timestral compare to TimesFM and Chronos?: Mistral reports that Timestral-209M beats TimesFM-3 and Falcon-2.0 while using fewer parameters, and that Timestral-28M beats Chronos-2 on GIFT-Eval and TIME. On the Hugging Face model card, Timestral-209M scores a MASE of 0.658 and CRPS of 0.446 on GIFT-Eval, where lower is better.

Try it

pip install git+https://github.com/mistralai/timestral.git