Timestral: forecasts without a giant model

Transcript

The original narration, preserved as published. Relative dates refer to the episode’s coverage window.

Forecasting does not have to mean asking a giant chatbot to guess a number. Today, October fifteenth, Mistral released Timestral: four time series models, ranging from five million to two hundred nine million parameters.

The models predict related numerical streams together, and output twenty one quantiles rather than just one answer. That gives a team a forecast distribution: a way to see uncertainty around demand, energy use, or sensor readings.

They can also use inputs known ahead of time, such as a production schedule or planned shutdown. That is different from leaking the answer into a test. The model separates historical inputs, future known inputs, and the values it must predict.

Mistral reports leading zero shot results on three forecasting benchmarks. Its published inference code runs on CPUs or GPUs. Smaller models offer another speed and accuracy tradeoff. Those are vendor results and capabilities, not a performance test we ran.

The code is Apache two point zero; the weights have a separate modified M I T license. The docs mark the model as public preview. Before relying on it, backtest against your existing forecast on held out historical data. A smaller model is useful when it predicts your problem well.