Mistral has launched Timestral, a family of time-series foundation models with its largest version, a 209M-parameter model, ranking first on all three main public forecasting benchmarks: GIFT-Eval, TIME, and fev-bench. The model offers state-of-the-art probabilistic, multivariate forecasts and is available in four sizes, balancing accuracy and efficiency.
TL;DR
- Mistral introduces Timestral, a family of time-series foundation models with sizes ranging from 5M to 209M parameters.
- The largest model, with 209M parameters, outperforms larger models on key benchmarks, offering state-of-the-art probabilistic, multivariate forecasts.
- Timestral is designed to be efficient, running on CPUs and excelling in zero-shot forecasting for diverse domains.
What happened
Mistral has unveiled Timestral, its first family of time-series foundation models. The largest model in the family, with 209M parameters, has achieved top rankings on the GIFT-Eval, TIME, and fev-bench leaderboards, outperforming larger models with fewer parameters. The model family includes four sizes: 5M, 28M, 123M, and 209M parameters, each optimized for a balance of accuracy and efficiency.
Timestral delivers probabilistic, multivariate forecasts and supports past-only and known-future covariates. The models are designed to run efficiently on CPUs, making them accessible for a wide range of applications. The family of models sits on the Pareto front of forecasting accuracy versus model size, delivering leading performance at each size.
Why it matters
For developers and startups, Timestral offers a powerful tool for time-series forecasting, which is crucial for applications in capacity planning, energy, supply chains, risk, and maintenance. The model's efficiency and zero-shot forecasting capabilities make it versatile for diverse domains, even where data is scarce.
Investors may see this as a strategic move by Mistral to expand its portfolio and compete in the growing market for AI-driven forecasting solutions. The model's performance on key benchmarks highlights its potential to drive innovation and solve complex forecasting challenges.
Key facts
- Timestral is available in four sizes: 5M, 28M, 123M, and 209M parameters.
- The 209M-parameter model ranks first on GIFT-Eval, TIME, and fev-bench leaderboards.
- Timestral supports past-only and known-future covariates, offering probabilistic, multivariate forecasts.
- The models are designed to run efficiently on CPUs, with inference times ranging from 10 ms to 140 ms depending on the model size.
- Timestral combines synthetic pre-training, covariate-aware attention, and long-context learning for improved forecasting.
- The model is available through Mistral Forge, allowing for fine-tuning on proprietary data and domain adaptation.
- Mistral has published a GitHub repository for users to run Timestral on their own data.
- The technical report with full results will follow the release.
Context
Time-series data is ubiquitous in enterprise settings, from sensor telemetry to observability metrics. Forecasts built on this data drive critical decisions in various industries, including capacity planning, energy, supply chains, risk, and maintenance. Traditional models like ARIMA and linear regression have been well-established but lack the ability to learn nonlinear patterns across many related series.
Deep learning models like PatchTST have added the capacity to learn nonlinear patterns, but they are typically trained from scratch on each dataset, limiting their transferability. Time-series foundation models, like Timestral, bring the pre-training approach to time series, learning patterns that transfer across series and forecasting zero-shot on data they have never seen.
Mistral's release of Timestral positions the company as a key player in the AI-driven forecasting solutions market. The model's performance on key benchmarks and its efficiency make it a compelling option for developers, startups, and investors looking to leverage advanced forecasting capabilities.
