Mistral AI launches Timestral, a CPU-based time series model family

Mistral AI has released Timestral, its first family of time series foundation models, according to MarkTechPost. The release includes four zero-shot forecasting checkpoints sized from 5 million to 209 million parameters. This launch marks a shift toward lightweight, CPU-capable infrastructure for advanced predictive analytics tasks.

The 209-million-parameter model ranks first among pretrained, zero-shot models on GIFT-Eval, TIME, and fev-bench benchmarks. It achieves a SQL skill score of 50.39 on fev-bench, ahead of Google’s TimesFM-3 at 48.99. The model runs inference on standard server CPUs without GPU acceleration, with weights totaling 837 MB in fp32 format.

Technical Architecture and Training

Timestral uses a causal, patch-based transformer with time and variate attention mechanisms. The largest checkpoint employs 24 layers and an embedding dimension of 768, processing patches of 32 time steps. It forecasts up to 1,024 steps in a single forward pass, extending horizons by feeding median outputs back as context.

Decoupled Horizon Forecasting allows learned tokens to read full history while keeping observed values in context. Covariate-aware attention prevents future leakage by directing information flow only from covariates to targets. Training involved synthetic data with trends and seasonality, followed by post-training on mixed data with longer contexts.

Developer Impact and Performance

Developers can use Timestral for probabilistic forecasts on standard hardware, accessing open weights and inference code. The models output 21 quantiles per time step and handle multivariate targets jointly. They accept past-only, known-future, and categorical covariates, transferring patterns across domains rather than training from scratch.

Unlike classical tools like ARIMA or deep models like PatchTST, Timestral leverages pretrained patterns for zero-shot forecasting. The 5-million-parameter model provides a lightweight option, while the 209-million-parameter version offers higher accuracy. All reported scores derive from single checkpoints without ensembling.

What should you do?

Security and technology professionals should evaluate Timestral for low-latency forecasting needs where GPU resources are constrained. Teams handling sensitive time series data should review the open weights and inference code for compliance with internal data governance policies. Organizations should test the model’s accuracy on their specific datasets before production deployment.

What to do and how to stay safe: Mistral AI

• Review the open weights and inference code for security vulnerabilities before integrating Timestral into production systems.
• Test the model on internal datasets to verify accuracy and ensure it meets specific forecasting requirements.
• Monitor for updates or patches from Mistral AI that may address any discovered issues in the initial release.
• Assess the impact of running inference on standard CPUs versus GPUs for cost and performance trade-offs.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is Timestral?

Timestral is a family of four time series foundation models released by Mistral AI, ranging from 5 million to 209 million parameters.

How does Timestral compare to Google’s TimesFM-3?

The 209-million-parameter Timestral model outperforms TimesFM-3 on several benchmarks, including a fev-bench SQL skill score of 50.39 versus 48.99.

Does Timestral require a GPU?

No, Timestral runs on standard server CPUs without requiring GPU acceleration, although GPU usage is optional.