theTNai
Get the brief

Models · Mistral AI · 15 Oct 2026

Mistral released Timestral, four small open time-series models for forecasting that run on a normal CPU.

On 15 Oct 2026 Mistral released Timestral, its first family of time-series foundation models, in sizes from 5M to 209M parameters. They forecast without training on the user's data first. Mistral says the largest model ranks first among pretrained zero-shot models on three public benchmarks.

What Mistral released

Mistral published Timestral on 15 Oct 2026. It is the company's first family of time-series foundation models: models that read a history of numbers, such as daily sales or power use, and predict what comes next. There are four sizes, with 5M, 28M, 123M and 209M parameters. All four share one design and one training method.

The models are zero-shot, which means they can forecast a series they have never seen without extra training. They return probabilistic forecasts: 21 quantiles rather than one line, so a user sees a range of likely outcomes. They accept several series at once, plus extra inputs such as past-only covariates and known-future covariates, for example a holiday calendar. Categorical covariates are supported. The context length is 32,768 time steps.

The weights use a modified MIT licence. Mistral did not state a price on its announcement page.

How it works and how fast it is

Mistral describes two ideas. The first, Decoupled Horizon Forecasting, keeps the full history in the sequence and uses learned forecast tokens that read it. The second, covariate-aware attention, gives each series a role (target, past-only covariate or known-future covariate), so information flows from covariates to targets only.

Most of the training data is synthetic. Mistral builds simple series with trends, seasonality and noise, then combines them into multi-series data with nonlinear links. Training has two stages: pre-training on purely synthetic data with a context of 8,192, then post-training on a mix of real and synthetic data with a context four times longer.

The models run on a CPU. Mistral's test used a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history. The 5M model took 10 ms and the 209M model took 140 ms.

The benchmark claims and what is still open

Mistral tested on GIFT-Eval, TIME and fev-bench. It says its 209M model ranks first among pretrained zero-shot models on all five metrics it reports, from a single checkpoint with no ensembling, and beats TimesFM-3 and Falcon-2.0 with fewer parameters. It says the 28M model beats Chronos-2, a model about four times larger, on GIFT-Eval and TIME. The 123M model comes within 0.2% of the 209M score at three quarters of its latency.

These are Mistral's own results, shown only as charts, so we cannot give exact scores. Mistral says it has submitted results to the GIFT-Eval and TIME leaderboards, and that a technical report will follow. Treat the ranking as a claim until the leaderboards confirm it.

For companies that want to adapt the models, Mistral Forge offers fine-tuning on private data, and deployment on-premises, in a private cloud or in air-gapped setups. Mistral notes that real production data can differ from benchmark data and shift over time.

How to try it

  1. Find the models in the Timestral collection on Hugging Face (huggingface.co/collections/mistralai/timestral).
  2. Get the code to run them on your own data from github.com/mistralai/timestral.
  3. Read the model page at docs.mistral.ai/models/timestral-1-0. Mistral's announcement does not describe API calls or a price.

How to get good results

Our practical advice, based on what the sources below say about it.

  1. Start with a small sizeThe 5M model forecasts in about 10 ms on a CPU, and Mistral says the 28M model beats a model about four times larger. Try a small one first on a laptop, and move up only if the forecasts are not good enough.
  2. Feed it known-future inputsThe models accept known-future covariates such as a holiday calendar. For Tamil Nadu sales or demand data, adding festival dates such as Pongal or Deepavali as known-future inputs is a natural thing to test.
  3. Use the range, not one numberTimestral returns 21 quantiles, not a single value. Plan stock or staffing around the range of likely outcomes, not a single line.
  4. Check on your own dataMistral says production data can differ from benchmark data and shift over time. Compare the forecast against a simple baseline on your own past data before relying on it.

Why it matters in Tamil Nadu

Tamil Nadu businesses, colleges and startups can try demand or energy forecasting on ordinary computers without a GPU.

Where this comes from

  • SourceMistral AI
  • Checked against the source19 Oct 2026
  • Written bytheTN.ai's AI agent, in its own words
  • Approved byPaul Jeevanesan, Founder and Chief EditorAlso founder and CEO of StudAI One, whose services wing is StudAI Works (related parties, see disclosures).

Spotted an error? Tell us: we fix confirmed errors within 24 hours and log them in public.

Cite this page

theTN.ai, "Mistral released Timestral, four small open time-series models for forecasting that run on a normal CPU.", 15 Oct 2026. https://thetn.ai/tech/2026-10-19-mistral-timestral/

More in Models