HeyHomees Lab Notes · No. 8 · Mistral · Published October 23, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Timestral-209M has 209M parameters and can handle up to 32,768 historical steps in one pass.
- It predicts 21 quantiles for horizons up to 1,024 steps, supporting univariate and multivariate data.
- On three major benchmarks it outperforms competing models, e.g., MASE 0.658 on GIFT-Eval.
What is it?
Timestral-209M is the largest member of the Timestral family of time-series foundation models, designed for zero-shot forecasting. It supports univariate and multivariate targets, past-only and known-future covariates, and outputs probabilistic forecasts across 21 quantiles.
How does it work?
The model uses a patch-based transformer architecture with time attention and variate attention, arranged in 24 layers and a 768-dimensional embedding. Input time steps are grouped into 32-step patches, each embedded together with missing-value indicators. Forecasts are produced in a single forward pass for up to 1,024 steps, with longer horizons generated block-by-block.
How well does it do?
On the GIFT-Eval benchmark Timestral-209M achieves a MASE of 0.658 and a CRPS of 0.446, beating t0-gamma and other listed models. On the TIME benchmark it records MASE 0.626 and CRPS 0.522, and on fev-bench it attains a SQL skill score of 50.39 and a MASE skill score of 39.49. All scores are reported by the authors.
What does it need to run?
Inference runs with PyTorch on either CPU or GPU. The model accepts up to 32,768 historical time steps (1,024 patches) and can forecast up to 1,024 steps (32 patches) per forward pass. No additional software requirements are listed beyond the provided GitHub code and environment setup.
By the numbers
- Parameter count: 209M
- Maximum context length: 32,768 time steps
- Quantiles output: 21 quantiles
- GIFT-Eval MASE: 0.658
The honest read
The reported benchmark numbers (e.g., MASE 0.658 on GIFT-Eval) are self-reported by Mistral and reflect zero-shot performance on the listed datasets; they do not include comparisons on other tasks or under different training regimes. The model’s ability to handle missing values is limited to NaN markers, and longer forecast horizons beyond 1,024 steps require block-by-block processing, which may affect latency. No external third-party evaluations are provided in the source.
What we don't know yet
- Training data and pre-training procedure are not described.
- Exact inference latency or memory footprint on typical hardware is not given.
Sources
- Model card, Hugging Face (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.