Mistral's Timestral tops GIFT-Eval with 209M parameters

Timestral arrives as enterprises increasingly rely on time-series data — from sensor telemetry to observability metrics — to drive decisions in capacity planning, energy, supply chains, risk and maintenance. Classical tools like ARIMA and linear regression learn from a single series' history, while deep-learning models such as PatchTST are typically trained from scratch on each dataset, so what they learn does not carry over to new domains. Time-series foundation models aim to change that by pre-training on a large and diverse corpus so patterns transfer across series, enabling zero-shot forecasts even where data is scarce.

Mistral's technical approach combines synthetic pre-training, covariate-aware attention and long-context learning. Most of the training data was synthetic, generated to capture trends, seasonality and diverse noise, then used to build multivariate series with nonlinear dependencies. The model supports past-only and known-future covariates — such as production schedules or weather forecasts — and outputs probabilistic forecasts across 21 quantiles, giving teams a sense of uncertainty rather than a single number. A decoupled horizon forecasting method keeps the full history in sequence while learned forecast tokens read it, so observed values serve as both target and context.

The release positions Timestral on the Pareto front of accuracy versus model size: for each size, no public model forecasts better, according to Mistral. Its 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. Beyond benchmarks, Mistral is offering the model through Mistral Forge for fine-tuning on proprietary data and deployment on clients' hardware, including air-gapped environments — a sign the company is targeting organizations that need to adapt forecasts to their own shifting data.