Mistral released Timestral, its first family of time-series foundation models in four sizes (5M, 28M, 123M, 209M parameters), which it says ranks first on the GIFT-Eval, TIME and fev-bench benchmarks for zero-shot probabilistic multivariate forecasting. The models support past-only and known-future covariates, run on CPUs without a GPU, and are available under a Modified MIT licence with a GitHub repository, plus fine-tuning and deployment options (including air-gapped) via Mistral Forge.
Read the full story on mistral.ai
What to doTry Timestral via the GitHub repository if you need CPU-based time-series forecasting, or use Mistral Forge for domain fine-tuning.
- Provider
- Mistral
- Source
- mistral.ai