Latest in AI
Mistral Timestral: Forecasting Models, CPU Inference and License Limits
Mistral’s new time-series models can run on a CPU, but the deployment decision starts with the data and the weights license.
Mistral released Timestral on October 15, 2026, a family of models for forecasting numerical data over time. The four models predict multiple related series and express uncertainty around those predictions. Downloadable weights and CPU inference make this worth examining for teams building demand, capacity or sensor forecasting tools.
The launch announcement describes models with 5 million, 28 million, 123 million and 209 million parameters. These are forecasting models rather than conversational assistants. The practical question is whether they improve a recurring prediction enough to change an operational decision.
What Timestral predicts
Timestral supports a history of up to 32,768 time steps and produces 21 quantiles at each future step. A quantile describes a position in the predicted distribution, giving a team more information than one expected value.
For inventory planning, the median forecast and a higher-demand estimate serve different purposes. A buyer worried about stockouts may care more about the upper part of the distribution. Whether those estimates are reliable still needs testing on that buyer’s own products and periods.
Timestral can also use covariates, additional inputs that help explain the target. Some exist only in historical data. Others, such as a planned promotion or calendar event, can be known for the forecast period. Keeping those categories separate matters: a future value that would not actually be available when making the decision can make a backtest look misleadingly good.
Running Timestral locally
The official inference repository supports CPU and GPU execution. Its Python interface loads a checkpoint through TimestralForecaster and accepts numerical arrays or tensors, with missing observations represented by NaN.
| Model | FP32 weights download |
|---|---|
| Timestral-5M | 21 MB |
| Timestral-28M | 111 MB |
| Timestral-123M | 490 MB |
| Timestral-209M | 837 MB |
These are weight-file sizes, not complete runtime memory requirements. Inputs, intermediate calculations and the surrounding application also consume memory.
Forecasts beyond 1,024 steps proceed in blocks, feeding earlier predicted medians back into the context. Test the actual horizon you need: a short-horizon result may not carry over unchanged.
The weights license has a revenue restriction
The code and model weights use different licenses. The inference code is Apache 2.0. The Timestral weights license is Modified MIT and includes an important eligibility condition.
You cannot exercise any rights under the default license if your company’s or employer’s global consolidated monthly revenue exceeded $20 million, or the equivalent in another currency, in the preceding month. The restriction also covers derivatives, modifications and combined works based on the model, including versions supplied by third parties. Affected users can request a commercial license, which Mistral may grant at its discretion.
Check the terms before investing in integration. The revenue test applies to the company or employer, so a small project inside a larger business can still fall outside the default license.
What the benchmarks establish
Mistral reports that Timestral-209M ranked first among pretrained, zero-shot models on GIFT-Eval, TIME and fev-bench at launch. Those vendor-reported results need checking against an existing forecasting system on the data it actually uses.
The useful next step is a time-ordered comparison with your current baseline. Test multiple forecast origins, inspect uncertainty calibration and measure the error that matters to the business. An overestimate and an underestimate can have very different costs.
CPU inference makes experimentation easier; it does not remove evaluation work. The more interesting opportunity is putting forecasts inside ordinary operational software, where the prediction can be checked against a decision and its eventual outcome.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn model releases and complex product decisions into clear stories developers and buyers can use.