Timestral is Mistral’s first family of time-series forecasting models, with four sizes ranging from 5M to 209M parameters. It supports multivariate forecasting and represents uncertainty ranges through quantiles.
The 209M-parameter model ranks first on GIFT-Eval, TIME, and fev-bench using a single checkpoint, without requiring an ensemble. Another highlight is that it can run inference on a CPU, making it suitable for monitoring pipelines that do not want to depend on GPUs.
Visualizing How Timestral Works
Imagine a factory sending vibration, temperature, and pressure data from a pump simultaneously. Timestral captures the relationships among these sensors, along with machine state and production schedules known in advance, to generate forecasts from multiple variables.
The output is not just a single number, but quantiles that represent a range of uncertainty. Teams can therefore see how wide the risks are in the upcoming trend and set alerts or plan maintenance more appropriately for the situation.
When Sensor Forecasting Needs to Look Beyond a Single Value
In a factory, teams need to monitor sensors from many machines to plan production capacity, detect early signs of equipment degradation, and prepare spare parts in time. This data is also connected to production schedules and planned downtime known in advance.
The problem is that forecasting each series separately only considers the history of that individual sensor, even though a pump’s vibration, temperature, and pressure may change together. Timestral is therefore designed to consider multiple series jointly while using related variables, allowing the forecast to better reflect the overall system.
Where Timestral Fits in Mistral’s Product Portfolio
Timestral is Mistral’s first family of time-series foundation models designed specifically for forecasting time-based data. Its key strength is that it is a numerical model that can be deployed in existing monitoring pipelines, working with sensor data and metrics already in place.
It can also be paired with language models in multimodal workflows. For example, Timestral can detect anomalies and then pass them to an agent that analyzes the causes using maintenance records, manuals, and related data. This allows Timestral to serve as the numerical forecasting component within a larger AI system.
Traditional models need to be retrained for each dataset, while Timestral attempts to transfer patterns from diverse data to previously unseen series, making it suitable for specialized or low-data tasks.
| Factor | ARIMA / linear regression | PatchTST | Timestral |
|---|---|---|---|
| Learning | Fits a single series | Retrained for each dataset | Pretrained on diverse data |
| Use with new data | Relies on the history of that series | Learned knowledge does not transfer to new domains | Supports zero-shot forecasting |
| Data patterns | Usually looks at one series at a time | Learns nonlinear patterns across multiple series | Transfers patterns across series |
How Much Do These Strengths Matter in Real-World Situations?
For a factory pump, Timestral can consider vibration, temperature, and pressure together, allowing it to capture relationships among the signals instead of forecasting each one separately.
Probabilistic forecasting shows the uncertainty range to the team, making it possible to configure alerts and plan maintenance based on risk rather than relying on a single predicted value.
For categorical covariates, machine states such as operating, idle, or under maintenance can be passed directly into the model, so there is no need to force these states into numerical representations.
Known-future covariates help the model account for production schedules, planned downtime, and weather forecasts. When the load is about to increase, the system can predict the resulting changes instead of treating them as abnormal events.
How Timestral Compares with Earlier Models and Competitors
Timestral ranges from 5M to 209M parameters, allowing users to choose a model based on the task. The 28M model also outperforms Chronos-2, which is approximately four times larger, on GIFT-Eval and TIME.
| Factor | Timestral | Competitor |
|---|---|---|
| Model size | 5M, 28M, 123M, 209M | Chronos-2 is approximately four times larger than Timestral 28M |
| GIFT-Eval and TIME | Timestral 28M performs better than Chronos-2 | Chronos-2 |
| Number of parameters | Timestral 209M has fewer parameters than TimesFM-3 and Falcon-2.0 | TimesFM-3 and Falcon-2.0 |
| Accuracy relative to size | All models are on the Pareto front | No information in the cited sources |
Strengths Shown by the Data and Limitations You Need to Read Completely
Timestral stands out by achieving the strongest results across three major benchmarks while supporting probabilistic and multivariate forecasts, allowing users to see both predictions and uncertainty. It is well suited to sensor data, monitoring systems, and planning tasks that require multiple signals to be considered together.
Pros
- +Achieves the strongest results on GIFT-Eval, TIME, and fev-bench
- +Supports both past and future covariates and runs on CPUs
Cons
- −Smaller models trade accuracy for speed
- −Domain-specific deployment may require fine-tuning through Mistral Forge
Usage Costs Are Not Measured by Parameter Count Alone
Timestral has four models ranging from 5M, 28M, 123M, to 209M parameters. The 5M model takes 10 ms for inference, while the 209M model takes 140 ms on an 8-thread CPU for a one-day forecast based on three weeks of historical data collected every 15 minutes. The 28M model is five times faster than the 209M model, while the 123M model scores within 0.2% and takes only three-quarters of the time of the largest model.
Accuracy therefore needs to be balanced against speed and cost for the actual task. The model also needs known-future covariates, such as production schedules or planned downtime, to provide context for the forecast. When integrated into a monitoring pipeline, teams must also design anomaly alerts and pass them to an agent for root-cause analysis using maintenance records, shift logs, and equipment documentation.
From Forecasting to Anomaly Detection and Analysis Systems
Timestral does not stop at predicting future values. It also serves as the numerical layer in a monitoring pipeline that continuously detects anomalies in sensor data.
When an abnormal signal is detected, an agent can compare it with historical sensor data to find similar failures and then pass it to a sub-agent for cause analysis. With context from maintenance records, work shifts, equipment manuals, and work orders, the analysis can be more complete.