Timestral

Time-series foundation models like Mistral's Timestral process input within a fixed context window; when the sequence length exceeds the window, the model simply discards the oldest tokens, all without any error, warning or logging.

Mistral released the Timestral time-series foundation model family on October 15, covering four sizes of 5M, 28M, 123M and 209M parameters; the largest model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench, with two-thirds the parameters of the previous best and without requiring ensembles.

Officials acknowledge that Timestral was primarily pretrained on synthetic data.

Timestral's 24×21 quantile outputs showed no zero crossings, i.e., quantile monotonicity held.

All of Timestral's median forecasts for the 24 months were below the actual values, with a constant deviation of -44.93.

In 24-month forecasts, Timestral's 80% quantile interval had an actual coverage of only 16.7%, at 25% for the first 12 months and dropping to 8.3% for the last 12 months.

The weight files on Hugging Face show that the 5M model corresponds to 5.21M parameters and the 209M model to 209.32M parameters.

Timestral supports 21-quantile probability forecasting and a context window of up to 32,768 steps.

The official claim is that Timestral ranks first on three public forecasting benchmarks.

The tester ran only the 5M and 28M models, not the 123M and 209M.

The release of Timestral means this track has moved from early exploration into a stage of multi-party competition, but the models have not yet formed a unified standard for pretraining data scale, architecture choice and evaluation benchmarks, so horizontal comparison remains difficult.

The original material did not disclose Timestral's model architecture, parameter count or benchmark results.

The 209M-parameter Timestral leads in the zero-shot model category on three benchmarks: GIFT-Eval, TIME and fev-bench.

The largest 209M-parameter version in the Timestral family can run inference on CPU.

The release information explicitly compares the Timestral 209M version with Google's TimesFM-3 and claims to beat the latter.

Timestral has not disclosed architectural details, but judging from its emphasis on zero-shot and CPU inference, it is closer to a lightweight continuous-value regression approach.

Before selection, it is advisable to wait until Timestral's model weights, technical report or third-party reproduction results are public before making a judgment.

On a laptop GPU, a 128-step forecast in webTimesFM takes about 50–300 milliseconds.

Mistral says the 209M version of Timestral ranks first among zero-shot pretrained models on GIFT-Eval/TIME/fev-bench.

The Timestral models are zero-shot, so users can make predictions without training on their own data.