Timestral 1.0 209M
AI model · Forecasting / time series · Time series → Time series by Mistral AI
State-of-the-art probabilistic time series forecasting at 209M parameters.
A 209M parameter time series foundation model for zero-shot probabilistic forecasting. It is designed for businesses needing efficient, high-accuracy predictions on standard hardware with support for external covariates.
Model specs
- Model family
- Timestral
- Released
- 2025-09-09
- Input
- Text
- Output
- Text
- Weights
- Open weights
- Parameters
- 209M
- Capabilities
- Reasoning
Category: Predictive Analytics
- Domains
- Time series forecasting
- Parameters
- 209M
- Sizes and variants
- Timestral 1.0 209M, 123M, 28M, 5M
- License
- Modified MIT
- Released
- 2025-09-09
- Parameter Count
- 209M active parameters
- 209M parameter time series foundation model
- State-of-the-art probabilistic forecasting
- Zero-shot forecasting with covariate support
Overview
Timestral 1.0 209M is Mistral AI's specialized time series foundation model designed for high-performance forecasting. It provides businesses with state-of-the-art probabilistic forecasts and zero-shot capabilities, meaning it can predict future trends without needing to be trained on specific historical data first. The model supports covariates, allowing users to incorporate external factors into their analysis for more accurate predictions. It is part of a broader family of time series models ranging from 5M to 209M parameters, offering flexibility for various computational constraints.
Forecasting Capabilities
- Probabilistic Forecasting: The model provides state-of-the-art probabilistic forecasts for time series data.
- Zero-Shot Capabilities: It supports zero-shot forecasting, allowing it to predict future values without specific training on the target dataset.
- Covariate Support: The model includes support for covariates, enabling the inclusion of external variables that influence the time series.
Performance and Efficiency
- CPU Optimization: The model is engineered to run efficiently on standard CPU hardware.
- Scalable Architecture: Part of a scalable family of models ranging from 5M to 209M parameters to balance performance and resource usage.
Use cases
- Financial analysts use Timestral to generate probabilistic forecasts for market trends without retraining on new datasets.
- Supply chain managers incorporate external covariates like weather or holidays to predict inventory requirements.
- Engineers run Timestral on standard CPU hardware to perform real-time forecasting for industrial sensor data.
Specifications
- Domains
- Time series forecasting
- Parameters
- 209M
- Sizes and variants
- Timestral 1.0 209M, 123M, 28M, 5M
- License
- Modified MIT
- Released
- 2025-09-09
- Parameter Count
- 209M active parameters
- Model Version
- 1.0
- License Type
- Modified MIT
- Primary Task
- Zero-shot time series forecasting
- Output Type
- Probabilistic forecasts
- Hardware Compatibility
- CPU and GPU optimized
FAQ
What is Timestral?
Timestral is a time series foundation model designed for zero-shot forecasting with probabilistic outputs and covariate support.
What model sizes are available for Timestral?
The Timestral family includes four sizes: 209M, 123M, 28M, and 5M parameters.
What license does Timestral use?
Timestral 1.0 models are released under a Modified MIT license.
Added by AntSpace from public sources (docs.mistral.ai).