← All posts
Guide

Mistral's Timestral: open forecasting models that run on a CPU, and how to test them on your own data

Mistral's Timestral forecasts sales, load or call volumes without a GPU. What it does, what Mistral claims, and a one-day test plan.

5 min read2 sources

Most AI news is about chat models. Last week Mistral released something many businesses can use more directly: Timestral, a family of open-weight forecasting models for time-series data such as daily sales, call volumes, power load or cash collections. Mistral says the models need no GPU to run and produce a range of likely values, not just a single number.

This guide covers what Timestral is, what Mistral claims for it, where it fits for Indian teams, and how to test it on your own data before you trust it.

What Mistral released

Mistral’s announcement, dated 15 October, describes Timestral as its first family of time-series foundation models. The main facts:

  • Four sizes: 5M, 28M, 123M and 209M parameters, all with the same architecture and training recipe. These are tiny next to language models.
  • Probabilistic output: each forecast comes as 21 quantiles. The Hugging Face model card for the 5M version gives the range as 0.01 to 0.99. You get the median as your main forecast plus bands such as the 10th–90th percentile, which tell you how uncertain it is.
  • Long history: up to 32,768 time steps of context. That is close to a year of 15-minute readings, or about 90 years of daily data.
  • Horizon: the model card says up to 1,024 steps per forward pass. Longer horizons are forecast block by block, with each block’s median fed back in as context.
  • Several series at once: inputs can be one or more target series, plus covariates that are either past-only (known only up to today) or known in advance (such as a holiday calendar or a planned price change). Missing values are marked as NaN.
  • Licence: Mistral calls it a “Modified MIT” licence. The model card says it adds a restriction against infringing third-party rights; read the LICENSE file in the repository before you ship.
  • Where it runs: on-premises, in a VPC or air-gapped. Weights are on Hugging Face and a GitHub repository is provided for running the models on your own data. Mistral Forge offers fine-tuning on proprietary data. The announcement gives no hosted API price.

One detail differs between the two pages. The announcement says categorical covariates are taken directly as categories, “rather than encoded as numbers”. The model card says categorical covariates are “target-encoded”. If you plan to use categorical inputs such as store type or region, check the repository code to see what actually happens.

What Mistral claims

All of these are Mistral’s own claims. The technical report with full results is not out yet (“will follow”, says the announcement).

  • The 209M model ranks first among pre-trained, zero-shot models on three public benchmarks, GIFT-Eval, TIME and fev-bench, from a single checkpoint without ensembling. Mistral says it submitted results to the GIFT-Eval and TIME leaderboards.
  • The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME. The 209M model beats TimesFM-3 and Falcon-2.0 while using fewer parameters.
  • Speed on a CPU: on a standard 8-thread server CPU, forecasting one day ahead from three weeks of 15-minute history took 10 ms with the 5M model and 140 ms with the 209M model. Mistral says the 28M model is five times faster than the 209M, and the 123M model scores within 0.2% of the 209M at three quarters of its latency.

“Zero-shot” means the model forecasts a series it has never seen, with no training on your data. That is the main attraction: you don’t need a data scientist to fit a model per product or per branch. It is also why you should test it, because public benchmarks may not look like your data.

Where it could fit for Indian teams

The use cases are ordinary forecasting jobs that many small and mid-sized companies still do in spreadsheets:

  • Retail and D2C: daily units per SKU or per outlet, with festival dates (Diwali, Eid, Onam, sale days) as known-future covariates.
  • Services and clinics: appointments or walk-ins per day, to plan staff.
  • Finance teams: daily collections or payment inflows, to plan cash.
  • Operations: call-centre volumes per hour, or electricity load from 15-minute meter data.

CPU inference and on-premises deployment matter here. You can run forecasts on the server you already have, and if your sales or customer data can’t leave your systems, you don’t need to send it to an outside API.

How to test it on your own data

Before Timestral goes into a planning process, check it against what you do today. A practical test that one person can run in a day:

  1. Pick three to five series that matter, such as top products or busiest branches. Use at least a year of history if you have it, so the model sees seasonality.
  2. Hold out the most recent period. For daily data, keep the last 4–8 weeks aside and forecast them as if they were the future. Repeat from several cut-off dates (a rolling backtest), not just one.
  3. Compare with a simple baseline. The usual one is “seasonal naive”: next Monday will look like last Monday. If Timestral can’t beat that clearly, it isn’t worth the change.
  4. Measure two things: error on the median (MASE or plain percentage error is fine), and whether the bands are honest. If you use the 10th–90th percentile band, about 80% of actual values should land inside it.
  5. Start with the 28M model. Mistral says it beats Chronos-2, a model about four times its size, and runs five times faster than the 209M. Move up to 123M or 209M only if the extra accuracy shows on your data.
  6. Then add covariates. Run the test again with the festival calendar or promotions as known-future inputs, and keep them only if they improve the backtest.

If the bands come out too narrow, or the model misses your specific seasonality, Mistral Forge’s fine-tuning is the next step. Check the cost of that against simply keeping your current method.

What to watch

  • The technical report. Until it’s published, the benchmark claims can’t be checked in detail.
  • The licence text. “Modified MIT” is not plain MIT. Read the added clause.
  • Your own backtest. A model that ranks first on public leaderboards can still lose to a seasonal baseline on a short, noisy series. Test before you trust it.

References

  1. Timestral: time-series foundation models — Mistral AI, accessed 2026-10-19
  2. mistralai/Timestral-5M (model card) — Hugging Face / Mistral AI, accessed 2026-10-19

This week in AI, every Sunday

One email with the releases that mattered and what they mean in practice. No spam.