Introducing Timestral

Summary

Timestral is Mistral's first family of time-series foundation models. They produce probabilistic, multivariate forecasts zero-shot, using past-only and known-future covariates. There are four sizes (5M, 28M, 123M, 209M parameters) under a Modified MIT licence, with weights on Hugging Face and code on GitHub.

Key points

• Context of 32,768 time steps; outputs 21 quantiles.
• The 209M model ranks first among pre-trained zero-shot models on GIFT-Eval, TIME, and fev-bench across all five metrics, according to Mistral. It uses a single checkpoint, with no ensembling. Exact scores appear only in charts.
• Mistral says the 28M model beats Chronos-2 (about four times larger) on GIFT-Eval and TIME, and the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.
• Runs on CPU: a one-day forecast from three weeks of 15-minute data takes about 10 ms (5M) to 140 ms (209M) on an 8-thread server CPU.
• Technique: "Decoupled Horizon Forecasting", covariate-aware attention, synthetic-first pre-training, then post-training on real and synthetic data at 4x longer context; pinball loss over the quantiles.
• Fine-tuning and custom training are offered through Mistral Forge, with on-premises, VPC, and air-gapped deployment. No API pricing is given.

Why it matters

Foundation models are spreading beyond text and images to numeric forecasting. Small open models that run on a CPU make zero-shot forecasting cheap enough to try before building a custom model. The benchmark claims are Mistral's own.