As of October 15, 2026, Mistral’s Timestral forecasting models are available for teams planning demand, capacity or equipment monitoring. The released software can run on a CPU without a dedicated GPU, giving operations teams another option for producing forecasts on their own infrastructure.
Our assessment: Timestral is worth evaluating if you have usable historical data, a specific planning decision and people who can integrate a model. It is not a ready-made inventory or staffing application. Mistral currently labels it Public Preview.
What forecast ranges could add to planning
Timestral can forecast related numerical series together and use information known in advance, such as a production schedule. Its output includes a middle estimate and multiple percentiles—values marking different points in the predicted range. The model card documents 21 such outputs.
For a planner, the useful question is whether those forecast ranges improve decisions about spare capacity or stock. An illustrative use would be checking whether an order plan still works when demand lands above the middle estimate. That is a proposed use, not a demonstrated Timestral business result.
The 10th-to-90th-percentile range represents a nominal central 80% interval: 90 minus 10. It does not mean the model has proved “80% accuracy”. Check how often later observations actually fall inside it, and whether the range is narrow enough to guide a decision. Prediction-interval principles.
Check the weights licence before a pilot
The source code is Apache-2.0 licensed. The downloadable model weights have separate Modified MIT terms. Those terms withhold rights if your company’s—or your employer’s—global consolidated revenue exceeded US$20 million, or its currency equivalent, in the preceding month.
That is a monthly revenue condition, not an annual threshold or a limit on your project’s budget. A small analytics team inside a larger employer therefore needs to check the organization’s eligibility before using the weights. The restriction also applies to derivatives and combined works.
The licence offers a route to request a separate commercial agreement, at Mistral’s discretion. Eligible users receive permission without a licence charge, subject to conditions including retaining attribution and permission notices. Download access alone does not settle permission for your organization. Full weights terms.
Which option belongs in the comparison?
Use the following as an evaluation shortlist, not a ranking. CPU support and uncertainty estimates are not unique to Timestral.
Option | Reason to include it | What would justify choosing it |
|---|---|---|
Your current method | It is the reference for the planning decision you already make. | Keep it if a new model adds no useful improvement after integration and operating costs. |
Timestral | Four sizes, from 5M to 209M parameters, with local CPU or GPU execution. Repository. | Better task results or a needed deployment capability, with usable forecast ranges and suitable licence terms. |
Chronos-2 | An existing alternative supporting CPU execution, related series and known-future inputs. Amazon’s model card. | A stronger fit on the same historical comparisons and operating constraints—not an assumed winner. |
Mistral reports leading Timestral results on three public forecasting benchmarks. It also reports 10 milliseconds for its 5M model and 140 milliseconds for its 209M model on an eight-thread server CPU, forecasting one day from three weeks of 15-minute observations. These are Mistral’s measurements, not independently reproduced results for this guide. The announcement says a full technical report will follow.
Those figures can justify a shortlist, but they do not establish accuracy on your products or the time needed to process a complete overnight planning run.
What evidence should decide the pilot?
The proposed evaluation below applies chronological forecasting evaluation to Timestral’s documented inputs. It has not been run for this article.
Match the real decision. If orders are placed four weeks ahead, evaluate that lead time. Choose historical forecast dates and use only information available at each date. Include your current method and a simple seasonal baseline, such as the previous week’s matching weekday for daily data with a weekly pattern.
Reconstruct the calendar honestly. For known-future inputs, use the promotion or production schedule that existed at each historical forecast date. A schedule revised afterwards—or actual weather supplied where only a forecast was available—would give the comparison information the planner did not have.
Check the range as well as the middle. Measure errors, persistent over- or underprediction, interval coverage and interval width by forecast horizon and relevant product or site group. Keep a final historical period untouched until model selection is complete.
Translate results into a planning choice. For stock, compare shortage and excess-stock consequences under stated cost assumptions; for capacity, compare unmet demand and spare capacity. Define acceptable tradeoffs before seeing results. A lower forecast-error score is not itself a measured saving.
Budget for a working forecast, not just model files
CPU compatibility removes the requirement for a GPU; it does not make deployment cost-free. Our suggested budget includes data preparation, memory, storage, complete-batch running time, monitoring and staff time. Record model and software versions so a successful comparison can be repeated.
No Timestral-specific hosted rate or commercial-licence price was established from the public pricing material checked. Do not substitute Mistral’s chatbot subscription or language-model token prices. The broader distinction between downloadable weights and a usable deployment also appears in our Step 5 self-hosting cost guide; that model has different hardware needs and licence terms.
The next step is a bounded historical comparison for one planning decision. Adopt Timestral only if the results and operating requirements support that choice. If the immediate need is a finished planning interface, or the historical records cannot support a fair comparison, solve that problem first.
Methodology: This decision guide uses published documentation, licence terms and forecasting-method references, checked October 15, 2026, with AI assistance. No model was run and no business outcome was measured. The announcement visibly displays October 15, but its structured metadata says July 15; the exact announcement timing remains unresolved. This guide uses an as-of date, not a claimed launch hour.