Mistral’s Timestral forecasts numbers, with a licence worth reading

More than one possible future

The four models have 5 million, 28 million, 123 million and 209 million parameters. They can forecast related series together and accept extra information, called covariates. A calendar of planned closures is known in advance; next week’s actual sales are not. A realistic test must preserve that distinction.

The model card describes 21 quantiles: predicted levels at different percentiles rather than one guaranteed answer. The median is the middle estimate; higher quantiles can help a planner consider a busier outcome. Whether those ranges match reality still needs checking on the organisation’s data. Machine-learning educator Damien Benveniste’s reaction makes that calibration question explicit; his post supplies an explanation, not an independently measured Timestral result.

Available code; early performance evidence

The inference repository runs on CPU or GPU and accepts up to 32,768 historical time steps. Longer forecast horizons are generated in successive blocks, feeding each block’s median predictions into the next. That is a useful limit to understand before treating a long forecast as one uninterrupted calculation.

Mistral reports leading pretrained, zero-shot results on GIFT-Eval, TIME and fev-bench. “Zero-shot” means applying the pretrained model without fitting it to each benchmark’s target series. Its documented inputs differ: TIME has no covariates, while the other tests can supply extra variables. These are Mistral’s reported comparisons; we have not verified an independent rerun. A fuller technical report is still forthcoming.

Aaron Abentheuer, who works on applied-AI design at Mistral, shared a launch-day interactive visualisation he helped create. It illustrates the forecast’s behaviour; it is not a customer outcome or an outside benchmark.

The modification in Modified MIT matters

The code uses Apache 2.0. The weights have a separate Modified MIT licence. Its condition 2 excludes use under that licence if the company’s—or employer’s—global consolidated monthly revenue exceeded $20 million in the preceding month. It directs those users to request a commercial licence from Mistral. That is a monthly company-revenue condition, not the price of a forecast or a threshold based on revenue earned from Timestral.

For teams considering demand planning, the next step is a historical backtest: use only information available at each past prediction date, compare against the existing forecast, and check whether the uncertainty ranges hold up. The release makes that investigation possible. It does not yet establish a sales uplift or a better inventory decision.