Mistral released four AI models on October 15 for forecasting numerical records, including sales, energy use and equipment readings, on a computer’s central processor, or CPU. Its biggest model produces the best overall forecasts in the company’s published comparison. SnowRock’s analysis of public scores for 100 forecasting tasks found that moving from its second-largest model to its largest reduced aggregate forecast error by about 0.4%.
The French company calls the family Timestral. The two largest versions have 123 million and 209 million parameters, the adjustable numbers that a model learns during training. Choosing the smaller version removes 41.1% of those parameters. It still wins on 39 of the 100 individual tasks when the comparison uses a measure that scores both predicted values and uncertainty.
For businesses, forecasting affects decisions about purchasing and equipment capacity. Amazon’s scientists described the underlying problem when they introduced their competing Chronos-2 model last year: demand depends on promotions, energy use depends on weather, and several measurements from the same computer can change together. Chronos-2 can receive those related readings and outside inputs together when generating predictions for a new forecasting assignment.
Mistral also attached a business condition to its downloadable model weights, the files containing those learned parameters. Companies whose consolidated monthly revenue exceeds $20 million fall outside the free license. That threshold makes company size relevant to the purchase decision, even though the software itself is small enough to run without a dedicated graphics processor.
Most of the accuracy arrives before the largest version
The public fev-bench result files provide a closer comparison than a leaderboard position alone. Timestral-209M had lower probabilistic forecasting loss on 61 tasks, while Timestral-123M had lower loss on 39. SnowRock matched task names, forecast lengths, evaluation windows and dataset identifiers before comparing the records. The analysis uses supplied forecasts’ scores; SnowRock did not rerun the models.
The 0.4% difference is calculated by comparing the two models’ errors on each task and taking the geometric mean of those ratios. This averaging method combines proportional changes across datasets with different units. The benchmark paper uses it for the same reason. A gain on an electricity dataset can therefore be combined with a gain on retail sales without treating kilowatts and units sold as interchangeable quantities.
Mistral’s published model table shows similarly small differences between its two largest versions on other tests. On GIFT-Eval, their point-forecast error scores are 0.659 and 0.658, respectively. Lower is better. On TIME, the scores are 0.628 and 0.626. Both versions also score ahead of Amazon’s Chronos-2 and Google’s TimesFM-3 in Mistral’s comparison of standalone forecasting models.
Four sizes, compared on the same 100 forecasting tasks
| Model | Parameters | Weight file | Relative forecast error |
|---|---|---|---|
| Timestral-5M | 5 million | 21 MB | 108.46 |
| Timestral-28M | 28 million | 111 MB | 103.97 |
| Timestral-123M | 123 million | 490 MB | 100.43 |
| Timestral-209M | 209 million | 837 MB | 100.00 |
SnowRock’s calculation from public fev-bench records. The largest model’s error is set to 100; lower is better. Each figure is the geometric mean of task-by-task probabilistic loss divided by Timestral-209M loss, multiplied by 100. Dataset fingerprints match across all four models. File sizes are 32-bit downloads, not total running memory.
Sources: Timestral model card, Mistral code repository. Error ratios calculated from public fev-bench result files; calculation saved in SnowRock’s reporting record.
Mistral’s CPU timing claim also favors the intermediate version. Its release announcement says the 123-million-parameter model takes three quarters of the largest model’s processing time. In a test forecasting one day from three weeks of 15-minute observations, the largest model took 140 milliseconds on a server CPU using eight threads. The smallest took 10 milliseconds. Those timings describe that test, rather than the time to process every series in a company’s records.
The code repository supports running downloaded models on either CPUs or graphics processors and loading previously saved files without another network request. Its published weight files range from 21 megabytes to 837 megabytes. Running memory also includes the forecasting software and input data. A business using this route operates the equipment itself, including the storage and software needed to turn company records into forecasts.
External inputs changed earlier forecasting results
One reason to include more information appears in fev-bench’s own research. On 42 tasks with changing external inputs, Amazon’s Chronos-2 reached a forecast skill score of 47.0% when it received those inputs, compared with 40.9% without them. Another model, TabPFN-TS, rose from 34.1% to 42.5%. The authors found more modest improvements from simply forecasting related series together. Their evaluation took place before Timestral’s release.
Timestral accepts both related series and additional inputs known before a forecast begins. Its documentation calls these inputs covariates. They can contain historical measurements, or future information already available, such as a holiday calendar or planned production schedule. The inference code’s description distinguishes those known inputs from the future values the model is supposed to predict. A scheduled shutdown can enter the forecast even though the equipment’s future readings remain unknown.
The test collections do not all ask the same questions. The TIME researchers assembled 50 fresh datasets and 98 forecasting tasks, in part to reduce dependence on repeatedly reused historical data. Their benchmark treats its measured variables as forecasting targets and excludes external covariates. Its results therefore assess a different use of the model from fev-bench tasks that supply outside information alongside the values being predicted.
TIME’s researchers also found that model rankings changed with the behavior of the records. In their original comparison, Chronos-2 led on series with stable statistical properties, while TimesFM-2.5 surpassed it when those properties shifted. Those findings predate Timestral. They document how an overall ranking can conceal different performance on stable demand, changing trends and other numerical patterns within a test collection.
Competitors offer different routes into a business
Mistral enters a field with established alternatives. Amazon introduced Chronos-2 on October 20, 2025, offering predictions across several related series and external inputs without additional training on each new forecasting task. Google described a BigQuery preview of TimesFM 3.0 in an October 12 announcement. That route puts forecasting inside Google’s database service, where customers can call it through database queries instead of operating a downloaded model.
Google illustrates its service with public New York Citi Bike records. Its documentation example predicts about 44,700 subscriber trips for June 1, 2016, with a 95% prediction interval from roughly 32,591 to 55,723. The exercise forecasts subscriber and casual trips together, using earlier trip durations and upcoming weekends as additional inputs. Its output includes an uncertainty range alongside the central estimate for each group.
Download rights differ from hosted-service access. Google’s TimesFM 3.0 weights license permits non-commercial, non-production use and says separate terms may apply to customers accessing its service. Mistral’s Modified MIT license permits broad reuse at or below its revenue cutoff. The cutoff uses the company’s or employer’s global consolidated revenue for the preceding month, rather than revenue earned by the particular project using Timestral.
Above $20 million in monthly revenue, Mistral says a company can request a separate commercial license, which it may grant at its discretion. The document gives no price for that agreement. Mistral’s forecasting code has a separate Apache 2.0 license. Downloading the code therefore does not remove the revenue condition attached to the model weights that the code loads.
A coding agent improves Google’s forecasts in another test
Mistral’s published table places Timestral first on GIFT-Eval, TIME and fev-bench among standalone forecasting models. A separate paper first posted October 14, TimeRSI, describes a coding agent that revises forecasting programs around an unchanged model. Its public result files carry different dataset fingerprints from Timestral’s, so those records do not establish a controlled head-to-head comparison. The paper instead measures improvements within its own forecasting system.
The TimeRSI team, including Haoxin Liu of Georgia Tech, tested changes against historical forecasting windows and selected programs using earlier errors. Their default system combined Google’s TimesFM 3.0 with a coding agent. The agent could change input transformations, selection of historical records and corrections to forecasts while leaving the model’s learned parameters fixed. In the paper’s default comparison, those changes raised the probabilistic forecast skill score from 48.984 to 51.243.
Mistral says its own models were trained first on synthetic numerical sequences, then on a mix of generated and real records. Its launch also offers adaptation through Mistral Forge, including further training on a company’s proprietary records. A full Timestral technical report describing the approach and complete results remained forthcoming as of October 16.
