From Mistral · · 1m 40s
Mistral launches Timestral: time-series foundation models
Mistral's new time-series foundation models forecast with 209M parameters, outperforming benchmarks on CPU.
In brief
Mistral's new time-series foundation models forecast with 209M parameters, outperforming benchmarks on CPU. Mistral released Timestral, a family of time-series foundation models that deliver probabilistic forecasts with 5M, 209M parameters, running efficiently on CPUs. Originally reported by Mistral.
Introducing Timestral
Mistral released Timestral, its first family of time-series foundation models delivering probabilistic forecasts with 209M parameters.
Benchmark Results
Timestral ranks first on GIFT-Eval, TIME, and fev-bench benchmarks with 209M parameters, outperforming previous state of the art.
Family of Models
Timestral offers four model sizes: 5M, 28M, 123M, and 209M parameters, balancing accuracy and efficiency for different use cases.
CPU Inference
Timestral runs efficiently on CPUs without GPUs, forecasting one day ahead from three weeks of 15-minute history in 10 ms with the 5M model.
Industrial Example
Timestral improves industrial sensor forecasts through multivariate, probabilistic, and categorical covariate inputs for better accuracy.
Key Technical Contributions
Timestral uses decoupled horizon forecasting, covariate-aware attention, synthetic-first data, and two-stage training for better time-series predictions.
Adaptation Options
Timestral supports zero-shot inference, domain-specific covariates, fine-tuning on proprietary data, and full control over training data.
What's Next
Mistral will publish a GitHub repository for running Timestral on user data and a technical report with full results.
Model Specifications
Timestral has 32,768 time steps context length, outputs probabilistic forecasts with 21 quantiles, and runs on-prem or air-gapped environments.
Why Timestral Matters
Timestral enables better forecasting for industrial plants, finance teams, and other enterprises by handling multivariate data and known future covariates.
Next Steps
Mistral will release a technical report and GitHub repository for Timestral, allowing users to run the models on their own data.
The big picture
Mistral released Timestral, a family of time-series foundation models that deliver probabilistic forecasts with 5M, 209M parameters, running efficiently on CPUs.
