On October 15, 2026, Mistral AI released Timestral, its first family of models for time-series forecasting. The models predict how measured values will change, such as the vibration and temperature of a pump or the load profile of a plant, and run on a standard server CPU. This applies to maintenance and plant management in plant engineering, operators of automated warehouses, and energy management at energy-intensive sites. However, whether a company is permitted to run Timestral at its own plant is determined by a revenue threshold specified in the license.
Timestral includes four models with 5 to 209 million parameters that, without training on proprietary data, generate forecasts with confidence intervals for multiple related time series. According to Mistral, a single server CPU is sufficient for this. The weights are licensed under a modified MIT license that excludes companies with more than $20 million in consolidated monthly revenue in the previous month. For such companies, options include a commercial license granted by Mistral at its discretion, Mistral’s hosted services, or a pre-trained model licensed under Apache 2.0, such as Chronos-2.
| Characteristic | Timestral | Chronos-2 | TimesFM 3.0 | Granite TTM r2 |
|---|---|---|---|---|
| Publisher | Mistral AI | Amazon | IBM | |
| Size | 5, 28, 123, and 209 million parameters | 120 million parameters | Not specified in the Model Card | Starting at 1 million parameters |
| Maximum History | 32,768 time steps | 8,192 time steps | Not specified in the Model Card | Depending on the version |
| Weight License | Modified MIT with a revenue cap | Apache 2.0 | TimesFM Non-Commercial License v1.0 | Apache 2.0 |
| In use in companies with monthly revenue exceeding $20 million | Only with a commercial license from Mistral | Permitted under the license | Only with a commercial license from Google | Permitted under the license |
What Mistral has released with Timestral
Timestral is a so-called foundation model for time series. It is primarily pre-trained on synthetic data and is designed to forecast time series it has never seen before, without requiring a separate model to be trained for each pump (Mistral AI, 15.10.2026).
The model reads up to 32,768 time steps of history; with 15-minute values, that is a little over 341 days, almost a year. It forecasts up to 1,024 steps in a single pass, about ten and a half days with 15-minute values. It processes related signals together, such as vibration, temperature and pressure of the same pump. And it takes known future values into account, for example the production schedule, planned shutdowns or a weather forecast. Instead of a single value, Timestral delivers 21 quantiles from which alarm thresholds with known uncertainty can be derived (Model Card Timestral-209M, 10/2026).
is how long the largest Timestral model takes, according to Mistral, on a server CPU with eight threads to produce a one-day forecast from three weeks of 15-minute values. The smallest model takes 10 ms.
Mistral AI, Timestral Announcement, October 2026
The revenue cap in the license
The program code on GitHub is licensed under Apache 2.0, the model weights are not. The license file for the weights permits use, modification and distribution like an MIT license, with one exception. According to the license text, anyone working for a company whose worldwide consolidated monthly revenue exceeded $20 million in the previous month may not exercise any rights under the license. According to the wording, this applies to all four model sizes and explicitly also to derivative, modified and combined works (License File Timestral-209M, 10/2026).
Spread evenly, $20 million per month corresponds to roughly $240 million in annual revenue. Many family-owned companies are below that, corporate groups and larger suppliers above it. Because the license looks at the previous month, a company with seasonal peaks can exceed the threshold in individual months. The text leaves open whether, for a subsidiary, its own revenue or that of the group counts. That question belongs with the legal department before anyone builds a pipeline on it.
Mistral names two routes for companies above the threshold: a commercial license, which Mistral grants at its sole discretion, or use via Mistral’s hosted services. As of October 17, 2026, Mistral has not published prices for either. The hosted route is a poor fit for plants whose machine data must not leave the company network.
Worked example: forecasts for 400 drives in an automated warehouse
An automated warehouse has 400 monitored drives, including stacker cranes, conveyors, and pumps. For each drive, three signals are received every 15 minutes: vibration, temperature, and current consumption. The maintenance department wants to know once an hour which drive will deviate from its normal operating pattern within the next 24 hours.
Based on the manufacturer's specification of 140 ms per daily forecast, 400 forecasts would theoretically require 56 seconds of computation time per run. Even if the computation time were ten times longer (for example, due to a longer history) it would still amount to about nine minutes per hour on a server without a graphics card. The hardware is rarely the bottleneck here.
In our experience, the bottleneck lies elsewhere. Timestamps from the control system and the historian don’t match, downtime isn’t flagged, and for very few drives is there documentation of when they actually failed. Without a failure history, it’s impossible to measure how much advance warning a model provides. Equally important is what happens after the alarm is triggered. Only when a warning is entered as a work order in the maintenance system and someone processes it does it prevent downtime.
For the load profile, the shift schedule as a known future value helps predict the quarter-hours in which a new annual peak load is likely, the value that determines the demand charge (see Grid Fees for 2027).
Your Options: Effort and Risks
- Stay with fixed thresholds. No project effort required. False alarms during load changes and delayed detection of creeping damage remain.
- Classic Models by Signal such as ARIMA or regression. License-free, but each time series requires its own adjustment, which is a never-ending task when dealing with hundreds of drives.
- Pre-trained model under Apache 2.0 such as Chronos-2 or Granite TTM, run on your own network. No licensing fees; a backtest can be completed in just a few weeks. The main risk is data quality.
- Timestral. It can be used freely below the revenue threshold and, according to Mistral, is currently the most accurate model in its class. Above the threshold, its use is contingent on a contract with Mistral, the price and terms of which are not disclosed.
For companies above the threshold, we recommend setting up data connectivity, backtesting, and alerting in such a way that the forecasting model is a replaceable component, and starting with a model licensed under Apache 2.0. If Timestral later demonstrates a measurable advantage with your data, it’s worth discussing a license with Mistral. On the other hand, those operating below the threshold should run Timestral in backtesting from the very beginning. The TimesFM 3.0 license permits only non-commercial testing; for production use, it requires a commercial license from Google.
Forecast backtest with your equipment data in three weeksWe collect 6 to 12 months of sensor or load profile data from up to 50 systems, run pre-trained models (and, where the license permits, Timestral) against your current threshold logic, and provide hit rates, false alarms, lead time, and hardware requirements. Everything runs on your infrastructure; the data never leaves your facility. Initial consultation: 30 minutes, free of charge, with a response within one business day.
Frequently Asked Questions
Can a corporation use Timestral for free?
According to the license terms, this is not the case if the global consolidated monthly revenue in the previous month exceeded 20 million U.S. dollars. In that case, the options are a commercial license or Mistral’s hosted services. The legal department will clarify how the threshold applies to subsidiaries.
Does Timestral need a graphics card?
Not according to Mistral. On a server CPU with eight threads, a one-day forecast based on three weeks of 15-minute data takes 10 ms with the smallest model and 140 ms with the largest.
Does Timestral replace a predictive maintenance system?
No. Timestral predicts measured values and highlights deviations. Predictive maintenance requires additional sensors, a data platform, documented failures, and a process for turning a warning into a maintenance order. Read more in the article Predictive Maintenance: What It Can and Cannot Do.
The Next Step
torck connects systems to the network, including older ones, and uses them to generate forecasts for ongoing operations, from sensor retrofits to a data platform with OPC UA and MQTT, all the way to integration with your maintenance system. During the initial consultation, we determine which systems are suitable for backtesting and which models are appropriate for your specific situation. Licensing reviews remain the responsibility of your legal department. More About Plant Monitoring. For load profiles and energy costs, Jouvoli is the right place to start.
Read more: Vibration Analysis of Bearings, Pumps, and Motors, From Sensor to Dashboard and Mistral Large 4 in your own AI operations.
This article outlines license terms to help you understand technical decisions. It is not legal advice. Please consult your legal department or a law firm to determine whether and how a license applies to your company.