Source: https://benchmarks.coralogix.com/timestral-benchmarks

# Timestral benchmarks on production telemetry.

Mistral released Timestral, a family of small forecasting models that read a stretch of history and predict what comes next. Models like this are usually measured on public datasets, which tells you something but not the thing we wanted to know: can one cope with a real production estate? So we tested Timestral performance against 19 series from the infrastructure that runs Coralogix, scored against four other foundation models and 12 classical and baseline methods.

The setup: a Coralogix account holding 1.2 petabytes and taking 40 terabytes a day, read through the cx CLI, with GLM 5.3 running the harness and Timestral producing the forecasts.

## How Timestral performed

Timestral was the most accurate forecaster we tested, and the fastest.

It ranked first a day and a week ahead against four other foundation models and 12 classical methods, with half the error of repeating yesterday. A Timestral size was the most accurate model on 12 of the 19 series, and on all 19 against the classical methods. The 28M beat every rival at both horizons and took 0.026 seconds a forecast on a laptop CPU, twice as fast as the quickest of them.

Hourly data from 25 June to 17 October 2026, with 12 to 16 weeks of history behind each forecast. We forecast every series 28 times a day ahead and 4 times a week ahead, against four other foundation models and 12 classical and baseline methods, and gave Timestral no configuration at any point. Dividing each series by its own average before scoring strips out the real volumes and leaves the shape.

Scores on this page use MASE: the forecast's error divided by the error you would get from simply repeating yesterday. Below 1 beats that guess, 0.5 is half its error, lower is better.

Nine of the series around the 10 October cutoff, forecast seven days ahead. Timestral followed the shape of most of them closely, held the daily and weekly rhythm a week out, and kept the real value inside its 80% band about 80% of the time. The ones it missed are in the edge cases below.

historywhat happenedTimestral forecast80% range

## The model

Timestral comes in four sizes, from 5 million to 209 million parameters. It takes a run of past values and returns a range of likely futures, with no per-series setup and no seasonality settings to choose. It runs on a laptop CPU in a fraction of a second, and the inference package installed and ran first time. Mistral's [announcement](https://mistral.ai/news/timestral), the [code](https://github.com/mistralai/timestral), and the [models](https://huggingface.co/collections/mistralai/timestral).

## Why we used production telemetry

Scoring well on a public forecasting dataset clears a low bar. Those datasets are tidy and well behaved, and a model released this year may have trained on them, so a good result tells you less than it appears to. Forecasting is worth having because it sits underneath capacity planning and predictive alerting, and both of those run against production.

So we used our own estate: 19 series over roughly four months. The data is private, which means Timestral cannot have seen any of it in training. It is also awkward in the ways real telemetry is always awkward.

- **The series behave nothing like each other.** Measured as variability against their own average, they run from 9% at the steadiest to 207% at the most erratic. A smooth storage curve, a spiky alert rate and a noisy p99 latency all have to be handled by the same model with the same settings.
- **Some have a working week in them and some do not.** Nothing labels which is which.
- **Collection has gaps,** and some series count upward and reset to zero.
- **Things happen that have never happened before.** Our final week contained two of them.

## Against other foundation models

We ran four other pretrained forecasting models on exactly the same 19 series, cutoffs and horizons: t0‑gamma, Chronos‑2, TiRex‑2 and TimesFM‑2.5. Every model got the identical raw history, gaps left as gaps, and returned the same nine quantiles. None of them was configured, and none can have seen this data in training.

- Series won a day ahead: 12/19
- Forecasts closer than TimesFM‑2.5: 67%
- Rivals the 28M outscores: 4 of 4

| Model | Maker | Size | Licence |
|---|---|---|---|
| Timestral | Mistral AI | 5M, 28M, 123M, 209M | Modified MIT |
| t0‑gamma | The Forecasting Company | Not stated | Apache 2.0 |
| Chronos‑2 | Amazon | 120M | Apache 2.0 |
| TiRex‑2 | NXAI | Not stated | Apache 2.0 |
| TimesFM‑2.5 | Google | 200M | Apache 2.0 |

We ran TimesFM‑2.5 because the weights of TimesFM‑3.0 are under a non-commercial licence.

### A day ahead, 532 forecasts

| Model | Score | 95% range | Inside 80% range |
|---|---|---|---|
| Timestral‑209M | 0.331 | 0.24–0.44 | 79.8% |
| Timestral‑123M | 0.333 | 0.24–0.44 | 81.1% |
| Timestral‑28M | 0.337 | 0.24–0.44 | 80.6% |
| t0‑gamma | 0.344 | 0.24–0.46 | 78.4% |
| Timestral‑5M | 0.358 | 0.27–0.46 | 80.6% |
| Chronos‑2 | 0.362 | 0.27–0.47 | 75.6% |
| TiRex‑2 | 0.363 | 0.26–0.47 | 82.0% |
| TimesFM‑2.5 | 0.379 | 0.28–0.49 | 81.8% |
| MSTL | 0.497 | 0.37–0.65 | – |

Timestral took the top three places. The 123M's error is 3% to 12% lower than each of the other models'. The margins are small next to the gap over the classical methods: every foundation model beats MSTL, and their 95% ranges all overlap.

### A week ahead, 76 forecasts

| Model | Score | 95% range | Inside 80% range |
|---|---|---|---|
| Timestral‑209M | 0.589 | 0.43–0.83 | 79.6% |
| Timestral‑123M | 0.590 | 0.43–0.83 | 79.6% |
| Timestral‑28M | 0.605 | 0.44–0.84 | 80.7% |
| t0‑gamma | 0.610 | 0.44–0.86 | 76.3% |
| Chronos‑2 | 0.615 | 0.44–0.88 | 76.7% |
| TiRex‑2 | 0.645 | 0.46–0.92 | 81.2% |
| Timestral‑5M | 0.647 | 0.47–0.90 | 80.0% |
| TimesFM‑2.5 | 0.648 | 0.47–0.92 | 78.6% |
| MSTL | 0.737 | 0.56–1.01 | – |

The order holds a week out, with t0‑gamma and Chronos‑2 closest behind. Mistral's model card ranks Timestral above t0‑gamma and Chronos‑2 on public benchmarks, and the same order appears on private data none of these models was trained on.

### How often the 123M model was closer, a day ahead

Timestral‑123M was closerthe other model was closer

1. TimesFM‑2.5**67** / 33
1. TiRex‑2**60** / 40
1. Chronos‑2**60** / 40
1. t0‑gamma**54** / 46

Out of 532 forecasts. Against TimesFM‑2.5, TiRex‑2 and Chronos‑2 the 95% range sits above 50%. Against t0‑gamma it runs from 49% to 59%, so that one is close to even. A week ahead the 123M was closer 58% to 64% of the time.

### Series by series

A day ahead, a Timestral size was the most accurate model on 12 of the 19 series. t0‑gamma won 3, Chronos‑2 2, and TimesFM‑2.5 and TiRex‑2 one each. A week ahead Timestral won 10, t0‑gamma 4 and Chronos‑2 2, with one each for TiRex‑2, MSTL and a straight line. The smaller sizes hold up too: the 28M scores better than all four rivals at both horizons, and the 5M is level with Chronos‑2 and TiRex‑2 a day ahead.

### Whose range means what it says

Every Timestral size kept the real value inside its 80% band between 79.6% and 81.1% of the time. Chronos‑2 and t0‑gamma came in under it, at 75.6% and 78.4% a day ahead, so their bands are a little narrow: the real value fell outside Chronos‑2's band 24% of the time. TiRex‑2 and TimesFM‑2.5 ran a little wide, at 82.0% and 81.8%.

### Time per forecast

| Model | Seconds |
|---|---|
| Timestral‑5M | 0.009 |
| Timestral‑28M | 0.026 |
| TiRex‑2 | 0.054 |
| Chronos‑2 | 0.060 |
| Timestral‑123M | 0.086 |
| t0‑gamma | 0.086 |
| Timestral‑209M | 0.120 |
| TimesFM‑2.5 | 0.348 |

The median of a separate timing run on the same machine. The 28M is twice as fast as the quickest rival and scores better than all four.

## Against classical statistics

MSTL, AutoTheta, CES, AutoARIMA and AutoETS, with naive baselines such as repeating yesterday, on the same forecasts.

- Series won a day ahead: 19/19
- Forecasts closer than MSTL: 75%
- Error below MSTL's, a day ahead: 33%

### A day ahead, 532 forecasts

| Method | Score | 95% range | Avg rank |
|---|---|---|---|
| Timestral‑209M | 0.331 | 0.24–0.44 | 3.44 |
| Timestral‑123M | 0.333 | 0.24–0.44 | 3.41 |
| Timestral‑28M | 0.337 | 0.24–0.44 | 3.63 |
| Timestral‑5M | 0.358 | 0.27–0.46 | 4.19 |
| MSTL | 0.497 | 0.37–0.65 | 6.27 |
| AutoTheta | 0.628 | 0.46–0.78 | 7.37 |
| CES | 0.644 | 0.50–0.80 | 7.45 |
| Repeat yesterday | 0.654 | 0.48–0.82 | 8.44 |
| AutoARIMA | 0.664 | 0.48–0.84 | 7.88 |
| AutoETS | 0.909 | 0.63–1.27 | 9.80 |

The top three sizes are tied: their ranges overlap completely, so we cannot separate them on this data. The best of the older methods, MSTL, lands at 0.497.

### How often the 123M model was closer, forecast by forecast

Timestral‑123M was closerthe other method was closer

1. AutoETS**94** / 6
1. Repeat yesterday**92** / 8
1. Repeat last value**90** / 10
1. AutoARIMA**87** / 13
1. AutoTheta**84** / 16
1. CES**84** / 16
1. MSTL**75** / 25

Out of 532 forecasts. The 95% ranges run from 91–97% against AutoETS down to 68–81% against MSTL.

### It found the working week on its own

On message-queue-bytes-in and message-queue-messages-in, Timestral predicted the weekend dip and the weekday peaks without being told the data has a weekly cycle. Methods that need that cycle declared up front were 4 to 13 times further out on those two series.

### A week ahead, 76 forecasts

| Method | Score | 95% range | Avg rank |
|---|---|---|---|
| Timestral‑209M | 0.589 | 0.43–0.82 | 3.34 |
| Timestral‑123M | 0.590 | 0.44–0.82 | 3.38 |
| Timestral‑28M | 0.605 | 0.44–0.84 | 3.45 |
| Timestral‑5M | 0.647 | 0.48–0.90 | 4.57 |
| MSTL | 0.737 | 0.55–1.00 | 5.25 |
| Repeat last week | 0.817 | 0.58–1.21 | 7.17 |
| AutoTheta | 1.045 | 0.80–1.35 | 7.80 |
| AutoARIMA | 1.052 | 0.78–1.38 | 8.03 |
| AutoETS | 1.653 | 1.24–2.27 | 10.08 |

AutoARIMA and AutoETS both pass 1.0 at this range, so a week out they do worse than repeating yesterday. Timestral holds at 0.590.

### Time per forecast

| Method | Seconds |
|---|---|
| Timestral‑5M | 0.010 |
| Timestral‑28M | 0.028 |
| CES | 0.077 |
| Timestral‑123M | 0.104 |
| Timestral‑209M | 0.155 |
| MSTL | 0.673 |
| AutoETS | 2.138 |
| AutoTheta | 5.385 |
| AutoARIMA | 46.012 |

AutoARIMA only saw the last 4 weeks of each series, to keep the run finishing at all, and still took 46 seconds per series. Timestral‑123M took a tenth of a second on the same machine.

## General findings

### The uncertainty range is honest

Timestral returns a band it expects to contain the real value 80% of the time. The real value landed inside that band between 79.6% and 81.1% of the time, at both a day and a week ahead. A band that means what it says is what makes forecast alerting workable: set a threshold at the edge of it and it fires about as often as you expect.

### Which size to run

The 28M, 123M and 209M models score the same within measurement error. Only the 5M trails, losing 61% of its head-to-head forecasts against the 123M. Since the 28M is four times faster than the 123M for the same accuracy, it is the one worth starting with.

## Edge cases

- **Anything without precedent.** A sharp drop in cache-memory-used and a growing backlog in message-queue-consumer-lag arrived in the final week, and every method missed both.
- **Smooth upward lines.** On a cleanly growing counter AutoARIMA and a straight line match or beat it. A week ahead, MSTL and a straight line took 3 of the 19 series between them.
- **Noisy series.** On p99 latency and load average every method is some distance out, and the gap between them narrows to very little.
- **Counters that reset.** A restart at zero drags the forecast down with it, costing one series 1.40% error against 0.21% once we trimmed the history. Feed rates or differences instead.

## Questions people ask about Timestral

How accurate is Timestral on production metrics?

On 19 production observability series it scored 0.333 MASE a day ahead, about half the error of repeating yesterday's value, and 0.590 a week ahead. It placed first against four other foundation models and 12 classical and baseline methods at both horizons. The best classical method, MSTL, scored 0.497 a day ahead.

How does Timestral compare with Chronos, TimesFM and other foundation models?

It ranked first a day and a week ahead against t0‑gamma, Chronos‑2, TiRex‑2 and TimesFM‑2.5 on the same production data. Timestral‑123M scored 0.333 a day ahead against 0.344 to 0.379 for the others, was closer on 54% to 67% of forecasts head to head, and a Timestral size was most accurate on 12 of the 19 series. The margins are small and the 95% ranges overlap.

Which Timestral model size should you use?

The 28M, 123M and 209M models score the same within measurement error, so the 28M is the practical choice at 0.028 seconds per forecast. Only the 5M trails, losing 61% of its head-to-head forecasts against the 123M.

How fast is Timestral on CPU?

Between 0.010 and 0.155 seconds per forecast depending on size, on a laptop CPU with no GPU. AutoARIMA took 46 seconds per series on the same machine, and that was with its history capped at 4 weeks.

Does Timestral need configuring for seasonal data?

No. It picked up the weekday peaks and weekend dip on our message queue series with no seasonality setting. Methods that need the cycle declared up front were 4 to 13 times further out on those series.

Are Timestral forecasts reproducible?

Yes. It produces its answer in a single pass with no sampling, so the same input gives bit-identical output. We reran 36 forecasts 10 times in fresh processes and every one matched exactly.

Where does Timestral struggle?

It cannot predict events with no precedent, it is matched or beaten by AutoARIMA and a straight line on cleanly growing counters, and its margin narrows to very little on noisy series such as p99 latency. Counters that reset to zero must be fed as rates or differences.

## Method and caveats

### Data

19 series at hourly resolution from 25 June to 17 October 2026, with 12 to 16 weeks of history behind each forecast. 28 starting points a day apart for the day-ahead task and 4 a week apart for the week-ahead task, none of them overlapping. Dividing each series by its own average removes the volumes and leaves the shape, which is all the scoring needs. Series names describe a role in general terms and nothing more should be read into them.

One figure on this page comes from an earlier warm-up run against the OpenTelemetry Astronomy Shop Demo rather than from production: the counter-reset example in the limits above. Everything else is measured on the production series.

### Scores

MASE divides the forecast's average error by the average error of repeating yesterday, measured over the history. Below 1 beats that baseline. We report the geometric mean across series, with MASE floored at 0.01, because one p99 series sits in the top bucket of its histogram and makes the naive forecast exactly right, which would otherwise leave the geometric mean undefined. WAPE, used for the counter example, is total error as a percentage of the total true value. Ranges are 95% confidence intervals from a cluster bootstrap over series, 2,000 resamples.

### The other methods

Classical models come from Nixtla statsforecast 2.1.1. Gaps were filled by linear interpolation for those models and passed through as missing to Timestral, which accepts them. AutoARIMA only saw the last 4 weeks of history, because fitting it on the full series did not finish in reasonable time. AutoETS chose non-seasonal models for these long hourly series and so forecasts a flat line beyond the first day. That is the library's own model selection rather than anything we configured.

The foundation models ran zero-shot with their published weights: t0‑gamma, Chronos‑2, TiRex‑2 and TimesFM‑2.5. Each received the same raw history as Timestral, missing values passed as missing, and returned nine quantiles from the 10th to the 90th percentile. Timestral's scores in that comparison are the same forecasts as in the rest of the page. On a quantile loss over those nine quantiles Timestral‑209M is first a day ahead, and a week ahead Chronos‑2 edges it by 8.511 to 8.518. Those scores come from a run on one NVIDIA A10G GPU. We repeated it on a laptop CPU, a MacBook, and six of the eight models gave identical scores to four decimal places, with TiRex‑2 moving 0.15% and TimesFM‑2.5 0.03%. The speeds on this page are from that CPU run, so Timestral's times there differ a little from those in the classical comparison.