Mistral AI Releases Timestral: Time Series Foundation Models

At a Glance
| Item | Value |
|---|---|
| Publisher | Mistral AI |
| Published | 2026-10-15 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Mistral AI has released “Timestral," the company’s first family of Time Series Foundation Models (TSFMs). Featuring support for probabilistic and multivariate forecasts, Timestral stands out by utilizing known-future covariates and categorical covariates in addition to historical data as inputs.
This model family is available in four parameter sizes: 5M, 28M, 123M, and 209M. Even the largest 209M model is reported to outperform previous state-of-the-art models with fewer parameters across all three major public time series benchmarks—GIFT-Eval, TIME, and fev-bench—taking the top spot among pre-trained zero-shot models. It is designed to run efficiently in CPU environments without requiring a GPU during inference.
Specifications
- Parameter sizes: 5M, 28M, 123M, 209M (four sizes available)
- Architecture: Common architecture shared across all four sizes (Decoupled Horizon Forecasting and covariate-aware attention mechanisms)
- Context length: 32,768 timesteps (large-scale pre-training on synthetic data at 8,192 steps, expanded 4x through a mixture of real and synthetic data)
- Input formats: Multivariate target series, past-only covariates, known-future covariates, categorical covariates
- Output format: Probabilistic forecasts (21 quantile forecasts using pinball loss as a discrete approximation of CRPS)
Performance
According to the publisher’s announcement, Timestral was evaluated on three major public time series forecasting benchmarks (GIFT-Eval, TIME, and fev-bench), securing the #1 spot across all five metrics for a single-checkpoint pre-trained zero-shot model without ensembles. The results have been submitted to the GIFT-Eval and TIME leaderboards.
Evaluation metrics and measurement conditions for each benchmark are reported as follows:
- GIFT-Eval (top 10 out of 83 comparisons, lower is better): GMean MASE, GMean WQL
- TIME (top 10 out of 35 registered on leaderboard, lower is better): GMean MASE, GMean CRPS
- fev-bench (top 10 out of 37 registered on leaderboard, higher is better): SQL skill score
Results regarding the trade-off between model size and accuracy (Pareto front) published by the creators show clear comparisons with existing large-scale time series models:
- The 28M model outperforms Chronos-2, which is approximately 4x its size, on both GIFT-Eval and TIME.
- The 209M model outperforms prior leading models such as TimesFM-3 and Falcon-2.0 with fewer parameters.
- Balancing within the family, the 123M model stays within 0.2% of the 209M model’s score while keeping latency at three-quarters (75%) of the 209M model.
- The 28M model runs 5x faster than the 209M model.
Regarding inference speed, measured values using a standard server CPU (8 threads) for predicting 1 day ahead from 3 weeks of 15-minute historical data are reported as 10 ms for the 5M model and 140 ms for the 209M model. While demonstrating efficient inference performance without requiring a GPU, it is also noted that there is a trade-off where smaller model sizes yield higher speeds at the cost of a certain decrease in prediction accuracy. All of these scores are figures published by Mistral AI itself.
Strengths and Use Cases
Timestral excels at zero-shot time series forecasting across diverse domains and at predicting and monitoring complex systems where multiple series influence each other.
The following tasks and operational setups are envisioned:
- Simultaneous multivariate time series forecasting:
Enables simultaneous modeling that accounts for dependencies without isolating sensor data that change in tandem rather than independently, such as vibration, temperature, and pressure in industrial plant pumps. - Probabilistic forecasting accounting for uncertainty:
Rather than a deterministic approach outputting a single predicted value, it can present a range of prediction uncertainty by forecasting 21 quantiles. This allows for setting alert thresholds and formulating maintenance plans based on risk ranges rather than relying on a single numerical value. - Direct utilization of future and categorical information:
Known future information (known-future covariates) known in advance—such as production schedules, planned outages, and weather forecasts—and categorical covariates like days of the week or equipment operational states (running, idle, maintenance) can be fed directly into the model. This improves forecasting accuracy without leaking future information from the target series. - Forecasting and planning in industrial equipment and financial domains:
In industrial settings, it is suited for drift detection prior to equipment failure, capacity planning across thousands of facilities, and spare parts management. In finance, it assists with position management, hedging, and treasury management through predictions of cash flows between accounts, transaction volumes, and volatility. - Autonomous agent workflows integrated with language models:
Beyond integration into existing monitoring pipelines, multimodal agent operations combined with language models are envisioned. For example, when Timestral detects an anomaly from a sensor data stream, an agent can compare it against embeddings of past failure histories, cross-reference inspection records, manuals, and shift logs to perform root cause analysis and present concrete diagnoses and next action proposals to on-site managers. - Domain adaptation and fine-tuning:
While delivering high performance even with zero-shot predictions from pre-trained models, it also supports fine-tuning tailored to proprietary sensor specifications or internal company data, generation of synthetic data, and expansion of domain-specific covariates via Mistral Forge.
How to Get It
Timestral’s code is available as a GitHub repository under the Modified MIT license. It supports deployment in on-premises environments, private clouds (VPCs), and air-gapped environments isolated from the internet.
It can be installed directly from the GitHub repository using the pip command in a Python environment (Python 3.10 or higher, PyTorch 2.4 or higher):
pip install git+https://github.com/mistralai/timestral.git
When using the uv package manager, you can also clone the repository and build the environment from the fixed dependencies (uv.lock):
git clone https://github.com/mistralai/timestral.git && cd timestral
uv sync --frozen
source.venv/bin/activate
Related Articles
- Mistral AI Releases Mistral Large 4 Open-Weight Model
- Mistral AI Announces Mistral Large 4 with 1.05T Parameters
- Mistral AI Announces Voxtral 3 Audio Models and Voice Agent SDK

