Timestral: Features, Benchmarks and Model Access

What is Timestral?

Timestral is a family of time-series foundation models released by Mistral on 15 October 2026. The models generate probabilistic, multivariate forecasts for enterprise data types such as sensor telemetry and financial metrics.

Unlike classical tools like ARIMA that train on a single data series, or deep learning models like PatchTST that train from scratch per dataset, Timestral uses pre-training across broad datasets. This design enables zero-shot forecasting on unfamiliar data without requiring prior domain training.

What are the main capabilities of Timestral?

Timestral supports multivariate, probabilistic forecasting along with continuous and categorical covariates. Its architecture incorporates Decoupled Horizon Forecasting, covariate-aware attention, synthetic pre-training, and long-context learning.

Instead of predicting variables in isolation, Timestral models related series together to capture interdependencies, such as correlations between pressure, temperature, and vibration. It outputs predicted quantiles to display forecast uncertainty for risk planning.

The model accepts direct categorical inputs, such as machine operational states or day-of-week categories. It also incorporates known-future drivers, such as production schedules or weather forecasts, to prevent expected events from triggering false risk alerts.

How does Timestral perform on benchmarks and speed?

Timestral achieved the top rank across the GIFT-Eval, TIME, and fev-bench public forecasting benchmarks among pre-trained zero-shot models. According to Mistral, its 209M parameter version outperforms larger competitor models using a single checkpoint without ensembling.

Timestral does not require GPUs for inference. On a standard 8-thread server CPU, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.

The smaller model variants balance speed against precision. The 28M model runs five times faster than the 209M model and beats Chronos-2 on GIFT-Eval and TIME, while the 123M model achieves performance within 0.2% of the 209M model score at three-quarters of the latency.

How to deploy Timestral in automated workflows

Timestral operates within standard data pipelines or combines with language models inside multi-modal diagnostic workflows. Organizations can implement an automated monitoring and analysis process using the following framework:

1. Run Timestral continuously on sensor data streams to monitor metrics and detect anomalies in real time.
2. Convert flagged anomaly signals into embeddings to compare them against historical failure datasets.
3. Trigger AI sub-agents to perform root cause analysis using maintenance records, operator notes, shift logs, and equipment manuals.
4. Generate a diagnostic summary and action plan based on current sensor predictions and retrieved context.

What are the limitations of Timestral?

Timestral is dedicated strictly to numerical time-series forecasting and does not generate conversational text responses on its own. Text-based diagnostic outputs require pairing the model with external language models.

While the smaller 5M and 28M parameter models offer higher execution speeds on CPU hardware, they sacrifice a small degree of forecast precision compared to the 209M parameter variant. Custom adaptation to highly specific enterprise domains requires fine-tuning through Mistral Forge.

Frequently asked questions

Sources: Mistral