Mistral AI has introduced Timestral, a foundation model family designed for time series forecasting. The architecture utilizes a causal, patch-based transformer with time and variate attention mechanisms. Each patch covers 32 time steps, and the model is capable of forecasting up to 1,024 steps in a single forward pass.
The flagship 209M parameter model features 24 layers and an embedding dimension of 768. It outputs 21 quantiles per time step and supports multivariate targets. The family also includes smaller versions with 5M, 28M, and 123M parameters, which are optimized for lower latency requirements.
Performance tests conducted on a standard server CPU with 8 threads show that the 5M model processes forecasts in 10 ms, while the 209M model requires 140 ms. The 28M model is reported to outperform Chronos-2 on the GIFT-Eval and TIME benchmarks despite being significantly smaller.
Timestral 1.0 209M is currently available in public preview. Users can perform fine-tuning via Mistral Forge, with deployment options including on-premises, VPC, or air-gapped environments. The package requires Python 3.10 and PyTorch 2.4 or newer versions.
