Mistral releases Timestral forecasting models that run without a GPU

Mistral is promising strong forecasts without a graphics card—or fresh training for every task. Released October 15, 2026, its Timestral family offers four downloadable models for predicting numerical trends. Mistral says the largest leads three public forecasting benchmarks, but that ranking has not been independently confirmed, and commercial access comes with a revenue-based restriction.

Small models, close benchmark margins

Timestral comes in versions with 5 million, 28 million, 123 million and 209 million parameters. These are time-series models: they predict how measurements recorded over time might continue, rather than generate chatbot replies. Potential uses include forecasting sales, electricity consumption and sensor readings. Zero-shot forecasting means using a pretrained model on new data without retraining it for that particular task.

Mistral’s model card puts Timestral-209M ahead of the listed competitors across GIFT-Eval, TIME and fev-bench. The comparison includes TimesFM-3, Chronos-2, Toto-2.0-2.5B and t0-gamma. On GIFT-Eval’s MASE error measure, where lower is better, Timestral scores 0.658 against t0-gamma’s 0.661. That is a lead of just 0.003, not a large separation.

The smaller checkpoints also make the size-versus-accuracy tradeoff visible. Mistral reports a GIFT-Eval MASE score of 0.659 for Timestral-123M, nearly matching the largest model’s 0.658. Those are company-run results, not an independent validation. Pasquale Pillitteri’s October 23 coverage says benchmark-maintainer confirmation is still pending; Mistral also cautions that business data can behave differently from public benchmark datasets.

The CPU claim has specific test conditions, too. Mistral measured a forecast of the next day’s 96 quarter-hour readings using three weeks of history, or roughly 2,000 observations. Its test used a server CPU with eight threads. The resulting timings describe that workload and hardware setup, rather than a guaranteed response time for every forecasting job.

10 millisecondsTimestral-5M

Mistral’s reported time for the smallest model on its one-day forecasting workload.

140 millisecondsTimestral-209M

Mistral’s reported time for the largest model under the same test conditions.

A forecast can use more than its own history

Timestral can process one series or several related series together. It also accepts supporting information, called covariates, alongside the measurements being predicted. The distinction between information available only in the past and inputs already known for the future is part of how the model handles a forecasting task.

• Targets are the measurements the model must predict. Future target values stay hidden while it builds the forecast.
• Past-only inputs supply supporting history, such as sales already recorded, without exposing future values.
• Known-future inputs can include a production plan or scheduled promotion. These remain visible across the forecast window.

The output is more than a single best guess. Timestral produces 21 quantiles, from 0.01 to 0.99: points across its predicted range of possible outcomes. Applications can use those outputs to express uncertainty around a forecast. The model also accepts missing observations, marked as NaN, rather than requiring every historical position to contain a measurement.

The largest checkpoint accepts up to 32,768 historical steps and predicts up to 1,024 future steps in one processing pass. Longer forecasts proceed in blocks: the model appends one block’s median prediction to its history before predicting the next. That extends the forecast beyond the single-pass limit, but the later blocks use generated predictions as part of their input.

Downloadable does not mean unrestricted

The weights are available through Hugging Face, with inference code on GitHub. Mistral’s installation instructions require Python 3.10 or later and PyTorch 2.4 or later, and support running on either CPU or GPU. The model card also links to code for reproducing the benchmark results, while listing the fuller technical report as still to come.

The weights and code have different licenses. The inference code uses Apache 2.0, while the weights carry a Modified MIT license. As Pillitteri’s licensing breakdown explains, free commercial use requires that the company—or the user’s employer—did not exceed $20 million in consolidated revenue in the previous month. The restriction also covers derivative and combined works.

Organizations above that threshold must seek a commercial license, which Mistral may refuse, or use Mistral-hosted services. Fine-tuning the downloadable model does not remove the revenue condition. The release therefore provides local access to the model, but not the same commercial permissions for every organization that downloads it.