{"id":5431,"date":"2026-10-19T11:29:10","date_gmt":"2026-10-19T08:29:10","guid":{"rendered":"https:\/\/taoailab.com\/zaman-serisi-tahmini-timestral\/"},"modified":"2026-10-19T13:18:59","modified_gmt":"2026-10-19T10:18:59","slug":"zaman-serisi-tahmini-timestral","status":"publish","type":"post","link":"https:\/\/taoailab.com\/en\/zaman-serisi-tahmini-timestral\/","title":{"rendered":"Mistral's new forecasting model runs on a standard server processor. What can a business do with it?"},"content":{"rendered":"
\n\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\n

Mistral's new forecasting model runs on a standard server processor. What can a business do with it?<\/h2>\n

\"Bir<\/p>\n

On October 15, Mistral released Timestral, its first family of time series forecasting models. They predict what comes next in data recorded over time, such as a machine's temperature, a store's daily sales or a building's electricity use. The smallest one forecasts a day ahead in about 10 milliseconds on a server processor, with no graphics card. Companies can download the weights and run them inside their own network. We think it deserves a close look, and large companies should read the license first, because it has a revenue limit.<\/em><\/p>\n\n\n\n
\n
10 ms<\/div>\n
ONE-DAY FORECAST, 5M MODEL, 8-THREAD CPU<\/div>\n<\/td>\n
\n
140 ms<\/div>\n
SAME TASK, LARGEST 209M MODEL<\/div>\n<\/td>\n
\n
32.768<\/div>\n
PAST TIME STEPS IT CAN READ<\/div>\n<\/td>\n
\n
$20M<\/div>\n
MONTHLY REVENUE CAP IN THE WEIGHTS LICENSE<\/div>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

First, what is time series forecasting?<\/h3>\n

A time series is any number recorded at regular intervals. A sensor reading a motor's vibration every second produces one, and so does a meter that logs a factory's power draw every 15 minutes. Many companies already collect several of these, often without calling them that.<\/p>\n

Time series forecasting uses that history to estimate the next values: next week's sales, tomorrow's peak load, or whether a pump is drifting toward a failure. The forecast helps with planning. When reality departs sharply from it, that gap is also a sign that something unusual is happening, which is a common basis for anomaly detection.<\/p>\n

In Mistral's speed test, the model forecasts one day ahead from three weeks of readings taken every 15 minutes. That is roughly 2,000 past values used to predict the next 96.<\/p>\n

What Mistral released<\/h3>\n

Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters. On disk they range from 21 MB to 837 MB, according to the GitHub page. The large language models behind today's chatbots are typically measured in billions of parameters, so these are small, single-purpose models.<\/p>\n

On a standard server processor with 8 threads, Mistral reports 10 ms for the smallest model and 140 ms for the largest on that one-day forecast. The 28M model is about five times faster than the 209M. The 123M comes within 0.2% of the 209M's score at about three quarters of its latency. Mistral says inference needs no GPU.<\/p>\n

Each forecast comes as 21 quantiles, from a cautious low to a cautious high, so the answer looks like \"between 410 and 470, most likely 440.\" That is easier to plan around than a single number. The model can forecast related series together, such as several sensors on one machine, and use how they move together. It also accepts information about the future that is already known, like a production schedule, a planned shutdown or a weather forecast, plus categories such as a machine's operating state. It reads up to 32,768 past time steps and forecasts up to 1,024 steps ahead in a single pass. Missing readings can be marked as missing.<\/p>\n

What zero-shot means here<\/h3>\n

Mistral calls Timestral a foundation model with zero-shot accuracy. It can produce a forecast for a series it has never seen, with no training on that series first.<\/p>\n

Classic forecasting methods are usually fitted to each data set before they can forecast it, which takes data preparation and an expert's time. With a zero-shot model you can get a first forecast on day one and judge whether it is good enough. If it comes close without being good enough, Mistral offers fine-tuning on a company's own data through its Forge platform.<\/p>\n

Most of the training data was synthetic, Mistral says. It was generated from simple processes and combined into multivariate series with nonlinear links between them. We think that is a sensible way to cover many patterns. How the model does on your own data still has to be tested separately.<\/p>\n

How Mistral says it compares<\/h3>\n

Mistral says the 209M model ranks first among pretrained zero-shot models on the field's three main public benchmarks: GIFT-Eval, TIME and fev-bench. It reports this from a single checkpoint, without combining several models. It also says the 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME.<\/p>\n

On GIFT-Eval, where lower error is better, the model card gives Timestral-209M a score of 0.658. Other models on the card include t0-gamma at 0.661, TimesFM-3 at 0.667 and Chronos-2 at 0.698. The lead over the closest rival is small, which means a buyer has several strong options and can choose on speed, license and fit.<\/p>\n

The example Mistral chose: a factory<\/h3>\n
\n

Mistral's announcement walks through plant monitoring. Sensor streams from equipment are forecast together, with the machine's state as a category and the production schedule as known future information. When readings depart from the forecast range, Timestral flags an anomaly.<\/p>\n<\/div>\n

Then an agent takes over. It looks for similar failures in the past and starts sub-agents that check maintenance records, logs, manuals and related sensor data to find a likely root cause. It then drafts a suggested diagnosis with its evidence and next steps.<\/p>\n

What to check before you plan around it<\/h3>\n

\"Kablolar<\/p>\n

Start with the license. The code is under Apache 2.0, but the model weights use a modified MIT license. It grants no rights to any company whose global consolidated monthly revenue exceeds $20 million, which is about $240 million a year. The restriction also covers derivatives and modified versions, which in our reading includes fine-tuned ones. Companies above the threshold can ask Mistral for a commercial license, which Mistral may grant at its discretion, or use the model through Mistral's hosted services. Your finance team can tell you quickly which side of the line your group is on. Ask before building anything.<\/p>\n

The benchmark results are Mistral's own. The technical report with full results has not been published yet; the announcement says it will follow. At the time of writing we found no independent evaluation of these results. Mistral says it has submitted them to the GIFT-Eval and TIME leaderboards.<\/p>\n

Timestral is also a preview. Mistral's model documentation labels Timestral 1.0 a Public Preview, and the GitHub repository does not accept outside code contributions. It answers issues on a best-effort basis.<\/p>\n

Finally, your own data decides. A benchmark mixes many domains, and your sales or sensors may behave differently. The simplest test is a backtest: hide the last few months of a series you already forecast, let Timestral predict them, and compare its error with your current method. If you have no current method, compare it with a naive forecast such as \"same as last week.\"<\/p>\n

What the model does not do for you<\/h3>\n

A forecasting model takes a clean table of numbers and returns a forecast. In our view, most of the work in a real project happens before and after that step. Someone has to get the readings out of machines, ERP and CRM systems, clean them and deal with the gaps. Someone has to decide which series matter, what counts as an anomaly and who gets alerted. The result then has to reach a dashboard or another system, and someone has to watch whether accuracy drifts as conditions change.<\/p>\n

\n

A small, fast, downloadable model makes the forecasting step cheaper and keeps the data in-house. For Turkish companies weighing KVKK, that is a real advantage. The data access, the alerting rules and the monitoring still have to be built, and we think they decide most of the result.<\/p>\n<\/div>\n

TAO AI LAB Perspective<\/h3>\n

\"TAO<\/p>\n

Timestral falls squarely within our area of work on prediction models built on big data. That area covers time series forecasting for demand, sales, load and traffic, anomaly detection for production defects and sensor failures, and risk scoring with early warning. Transformer-based time series models, the family Timestral belongs to, are among the methods we work with.<\/p>\n

The work around the model is the part we focus on: bringing in data from IoT devices, ERP, CRM and log systems, preparing it, and turning forecasts into dashboards and APIs that other systems can use, as well as monitoring models for drift and retraining them when needed. We have started working with offline models too, and a release like this one makes keeping data inside the company more practical. We are excited about it, and we look forward most to testing models like Timestral on real Turkish operational data.<\/p>\n

Frequently asked questions<\/h3>\n

What is Timestral?<\/strong><\/p>\n

\n

A family of four time series forecasting models from Mistral, released on October 15, 2026, in sizes from 5M to 209M parameters. It forecasts future values of data recorded over time and gives a range of likely outcomes.<\/p>\n<\/div>\n

Does Timestral need a GPU?<\/strong>
No. Mistral reports about 10 ms for the smallest model and 140 ms for the largest on a one-day forecast, using an 8-thread server processor.<\/p>\n

Can a large company use Timestral for free?<\/strong>
Not if its global consolidated monthly revenue exceeds $20 million. Above that threshold the weights license grants no rights. The company has to request a commercial license from Mistral or use the model through Mistral's hosted services.<\/p>\n

What is time series forecasting used for?<\/strong>
Planning and early warning: demand and sales forecasts, energy load planning, and spotting equipment problems when sensor readings depart from what was expected.<\/p>\n

If you could reliably forecast one number in your business a day ahead, which number would you pick? Tell us in the comments.<\/p>\n

Sources<\/h3>\n