Mistral launches Timestral time-series forecasting models from 5M to 209M parameters

Mistral AI has introduced Timestral, a family of compact time-series forecasting models ranging from 5 million to 209 million parameters. Unlike general-purpose language models, Timestral is designed to predict numerical sequences, quantify uncertainty and support operational decisions involving historical observations.

The family targets demand forecasting, financial planning, energy consumption and infrastructure monitoring. Mistral reports low-latency CPU inference, including approximately 140 milliseconds for a specific daily forecasting workload on eight CPU threads.

The central question is whether specialized foundation models can outperform established statistical and machine-learning approaches while remaining economical to deploy across thousands of series.

··········

TIMESTRAL AT A GLANCE: FOUR MODELS, 32K TIMESTEPS AND PROBABILISTIC FORECASTS.

Mistral announced Timestral on October 15, 2026, releasing four pretrained forecasting models with downloadable weights and a lightweight inference library.

The models are Timestral-5M, Timestral-28M, Timestral-123M and Timestral-209M.

........

Specification	Timestral
Developer	Mistral AI
Release date	October 15, 2026
Model variants	5M, 28M, 123M, 209M
Architecture	Patch-based transformer
Maximum context	32,768 timesteps
Maximum horizon per forward pass	1,024 timesteps
Patch size	32 timesteps
Forecast outputs	21 quantiles per timestep
Supported series	Univariate and multivariate
Additional inputs	Historical and known-future covariates
Zero-shot forecasting	Supported
Missing observations	Supported
CPU and GPU inference	Supported
Model weights	Downloadable
Model license	Modified MIT
Inference library	Apache 2.0
Domain adaptation	Mistral Forge

........

The 32,768-step context refers to numerical observations, not language tokens. For hourly data, this represents approximately 3.74 years; for 15-minute observations, approximately 341 days.

··········

THE FOUR TIMESTRAL MODELS BALANCE FORECASTING ACCURACY AND INFERENCE COST.

Mistral offers four sizes to let developers trade off accuracy, latency and memory requirements.

........

Model	Parameters	FP32 weight size	Main trade-off
Timestral-5M	5 million	21 MB	Lowest latency and memory
Timestral-28M	28 million	111 MB	Lightweight general forecasting
Timestral-123M	123 million	490 MB	Near-largest-model accuracy
Timestral-209M	209 million	837 MB	Highest reported accuracy

........

Checkpoint sizes are not total runtime memory requirements; activations, inputs and buffers also consume memory.

Mistral reports the 28M model is approximately five times faster than the 209M model under its benchmark configuration, while 123M approaches the largest variant's accuracy at lower latency.

A business forecasting thousands of predictable series may prefer a smaller checkpoint and reserve larger models for difficult tasks.

··········

HOW TIMESTRAL'S PATCH-BASED TRANSFORMER PROCESSES TIME SERIES.

Timestral groups consecutive numerical observations into patches of 32 timesteps.

A 32,768-observation input becomes 32,768 ÷ 32 = 1,024 patches, reducing the number of elements processed by the transformer.

Temporal attention learns relationships across time, while variate attention captures relationships among different measured variables.

For example, an electricity-demand forecast can consider temperature, industrial activity and regional consumption together.

Missing-value indicators help distinguish absent observations from actual numerical values.

This architecture supports both single-variable and multivariate forecasting without requiring a separate model for every series.

··········

MULTIVARIATE FORECASTING ALLOWS RELATED VARIABLES TO BE PREDICTED TOGETHER.

Traditional pipelines often fit independent models to each measured variable. Timestral can jointly process related series.

Regional electricity demand may share drivers such as weather and working days. Sales of related products may respond to the same promotions and holidays.

Joint forecasting can exploit these relationships, but adding irrelevant or noisy variables can introduce complexity.

Organizations should compare multivariate results with univariate baselines before assuming that additional inputs improve accuracy.

··········

PAST-ONLY AND KNOWN-FUTURE COVARIATES IMPROVE FORECASTING CONTEXT.

Covariates are explanatory variables used alongside the target series.

Past-only covariates are historically observed but not necessarily known in advance. Known-future covariates have values available for the forecast horizon.

........

Covariate	Example	Future value known?
Day of week	Calendar weekday	Yes
Public holiday	Holiday calendar	Yes
Scheduled promotion	Planned discount	Yes
Factory shutdown	Maintenance schedule	Yes
Realized customer demand	Future sales	No
Actual future temperature	Measured weather	No
Historical advertising spend	Recorded expenditure	Not necessarily

........

The distinction prevents data leakage: using information in evaluation that would not have been available when the forecast was made.

··········

TIMESTRAL GENERATES 21 QUANTILES INSTEAD OF A SINGLE POINT FORECAST.

Timestral produces probabilistic predictions across 21 quantiles, including the median and tail percentiles.

........

Illustrative quantile	Predicted units	Interpretation
10th percentile	820	Lower-demand scenario
50th percentile	1,000	Median forecast
90th percentile	1,240	Higher-demand scenario

........

These numbers are hypothetical, not model benchmark results.

If intervals are calibrated, approximately 80% of comparable outcomes should fall between the 10th and 90th percentiles over repeated forecasts.

Quantiles can inform inventory buffers, reserve capacity and cash-flow stress testing.

However, predicted intervals are not guarantees. Backtesting must establish whether uncertainty estimates are calibrated.

··········

THE 32,768-TIMESTEP CONTEXT WINDOW SUPPORTS LONG HISTORICAL PATTERNS.

The maximum historical context corresponds to different durations depending on sampling frequency.

........

Sampling frequency	Approximate history
Every minute	22.8 days
Every 15 minutes	341 days
Hourly	3.74 years
Daily	89.7 years
Weekly	630 years

........

These conversions describe capacity, not recommended history length.

Older observations may help capture seasonality but become misleading after structural changes, product launches or shifts in customer behavior.

More history does not automatically produce more accurate forecasts.

··········

FORECASTS CAN EXTEND BEYOND 1,024 TIMESTEPS THROUGH RECURSIVE GENERATION.

Timestral can predict up to 1,024 future steps in one forward pass.

Longer forecasts can be produced in successive blocks, using earlier predictions to extend the context.

Recursive generation introduces uncertainty because early errors can propagate into later predictions.

The ability to generate a long numerical sequence should not be confused with reliable long-term forecasting.

Organizations should measure forecast error separately across short, medium and long horizons.

··········

MISTRAL REPORTS LEADING RESULTS ON GIFT-EVAL, TIME AND FEV-BENCH.

Mistral evaluated Timestral across three forecasting benchmarks: GIFT-Eval, TIME and fev-bench.

The company reports that Timestral-209M achieves leading aggregate performance among selected pretrained and zero-shot models.

Its evaluation setup includes 97 GIFT-Eval configurations, 98 TIME tasks and 100 fev-bench tasks.

These results provide useful comparisons, but leaderboard performance does not guarantee superiority on every business dataset.

Data characteristics, evaluation horizons and benchmark methodology influence outcomes.

··········

TIMESTRAL BENCHMARK RESULTS SHOW DIMINISHING RETURNS AS MODEL SIZE INCREASES.

Mistral's reported results show accuracy improvements with model size, but smaller gains between the two largest checkpoints.

........

Model	GIFT MASE ↓	GIFT CRPS ↓	TIME MASE ↓	TIME CRPS ↓
Timestral-5M	0.705	0.479	0.652	0.545
Timestral-28M	0.678	0.461	0.638	0.534
Timestral-123M	0.659	0.448	0.628	0.525
Timestral-209M	0.658	0.446	0.626	0.522

........

MASE measures scaled point-forecast error; lower is better.

CRPS evaluates probabilistic forecast quality, combining accuracy and distributional calibration; lower is better.

GIFT-Eval MASE improves only from 0.659 to 0.658 between 123M and 209M, approximately 0.15% relative improvement.

The 123M checkpoint may therefore offer a favorable accuracy-latency compromise, although aggregate scores can conceal dataset-specific differences.

··········

FEV-BENCH RESULTS SUPPORT THE SAME ACCURACY-EFFICIENCY TRADE-OFF.

Mistral also reports the following fev-bench scores.

........

Model	SQL skill score ↑	MASE skill score ↑
Timestral-5M	46.19	34.54
Timestral-28M	48.42	37.27
Timestral-123M	50.18	39.19
Timestral-209M	50.39	39.49

........

Higher skill scores indicate better performance under the benchmark methodology.

The 123M-to-209M improvement in SQL skill score is only 0.21 points.

For large-scale deployments, this incremental accuracy benefit should be evaluated against additional latency and compute requirements.

··········

CPU INFERENCE REACHES 10 MS WITH 5M AND 140 MS WITH 209M.

Mistral reports CPU latency using eight threads on a server, forecasting one day from three weeks of 15-minute historical observations.

The test uses 2,016 historical observations and predicts 96 future steps.

Reported latency is approximately 10 milliseconds for Timestral-5M and 140 milliseconds for Timestral-209M.

These are specific benchmark conditions, not universal guarantees.

Latency varies with hardware, input length, number of series, covariates, concurrency and data-processing overhead.

The results nevertheless indicate that forecasting workloads may be practical without dedicated GPUs.

··········

WHAT CPU LATENCY MEANS FOR FORECASTING 100,000 TIME SERIES.

Suppose an organization refreshes 100,000 forecasts every 15 minutes, a 900-second interval.

At 10 milliseconds per forecast, the nominal serial workload is 1,000 seconds.

At 140 milliseconds, it is 14,000 seconds.

Dividing by the refresh interval gives approximately 1.11 and 15.56 equivalent continuously utilized processing slots, respectively.

These are simplified illustrative calculations, not actual CPU provisioning recommendations. The benchmark uses eight threads, and parallel scaling may be nonlinear.

They show why small latency differences matter when forecasting is repeated at scale.

··········

ZERO-SHOT FORECASTING REDUCES THE NEED TO TRAIN A MODEL FOR EVERY DATASET.

Timestral is designed to forecast unseen datasets using pretrained weights, without fitting a separate model for each series.

This can reduce maintenance costs for organizations with thousands of product, customer or sensor histories.

Zero-shot capability does not guarantee the best accuracy on every dataset.

A specialized statistical or supervised machine-learning model may outperform a general forecasting foundation model when sufficient domain-specific training data is available.

Comparisons should include seasonal naïve baselines and appropriate statistical alternatives.

··········

TIMESTRAL COMPETES WITH CHRONOS, TIMESFM AND OTHER FORECASTING FOUNDATION MODELS.

Mistral positions Timestral against forecasting foundation models such as Chronos, TimesFM and Falcon.

It reports that Timestral-28M outperforms a larger Chronos-2 model on selected benchmarks and that Timestral-209M performs strongly against TimesFM-3 and Falcon-2.0.

These are benchmark-specific claims, not evidence of universal superiority.

........

Comparison dimension	Why it matters
Forecast accuracy	Quality of predictions
Calibration	Reliability of uncertainty intervals
Zero-shot performance	Generalization to new datasets
Covariate support	Use of explanatory variables
Context length	Available historical information
Latency	Operational responsiveness
Model size	Hardware and memory requirements
License	Commercial rights and restrictions
Fine-tuning	Domain adaptation

........

Organizations should test candidate models on their own forecasting tasks and evaluate total operational value.

··········

FINANCIAL PLANNING APPLICATIONS INCLUDE REVENUE, CASH FLOW AND WORKING CAPITAL.

Timestral can support forecasting of revenue, payment volumes, operating expenses, liquidity and other financial time series.

Known-future covariates may include scheduled promotions, holidays or contractual events.

Future customer payments, however, cannot be treated as known in advance merely because a model can accept covariates.

Acquisitions, accounting changes and unusual transactions can produce structural breaks that reduce the usefulness of historical patterns.

The model can support financial planning but does not replace scenario analysis or financial judgment.

··········

PROBABILISTIC CASH-FLOW FORECASTS CAN SUPPORT LIQUIDITY RISK MANAGEMENT.

Consider a hypothetical 30-day cash-receipts forecast.

........

Quantile	Forecast receipts
P10	€1.6 million
P50	€2.0 million
P90	€2.5 million

........

If unavoidable payments total €1.8 million, the P10 scenario implies a potential €200,000 shortfall.

Such distributions can help determine liquidity buffers and contingency financing.

The values are illustrative. Actual decision quality depends on calibration, historical backtesting and the treatment of exceptional cash flows.

··········

INDUSTRIAL, ENERGY AND RETAIL APPLICATIONS CAN BENEFIT FROM LOW-COST FORECASTING.

Energy operators can forecast consumption, retailers can estimate product demand and manufacturers can predict throughput or sensor measurements.

These environments often generate large numbers of regularly sampled series.

CPU-friendly inference may reduce infrastructure costs and simplify deployment.

However, business impact depends on the type of error.

Underestimating demand can create stockouts or capacity shortages; overestimating demand can create excess inventory or unnecessary reserves.

Evaluation should reflect these asymmetric economic consequences.

··········

FINE-TUNING THROUGH MISTRAL FORGE SUPPORTS DOMAIN-SPECIFIC FORECASTING.

Mistral describes domain adaptation through Mistral Forge.

Fine-tuning may improve performance on specialized industrial, financial or operational data that differs from the model's general training distribution.

Evaluation must preserve chronological order. Random train-test splits can leak future information into historical training.

Backtesting should reproduce the information available at each historical forecast date.

Organizations should also compare the benefits of customization against its engineering and maintenance costs.

··········

ON-PREMISES AND AIR-GAPPED DEPLOYMENT SUPPORTS DATA-SENSITIVE APPLICATIONS.

Downloadable Timestral weights allow deployment in local, private-cloud and potentially air-gapped environments.

This can be useful for sensitive financial data, regulated operations and industrial systems.

Local inference can avoid transmitting inputs to external APIs, but organizations remain responsible for access controls, infrastructure security and monitoring.

Small checkpoint sizes lower hardware barriers without eliminating operational requirements.

··········

THE MODIFIED MIT LICENSE REQUIRES REVIEW BEFORE COMMERCIAL DEPLOYMENT.

Timestral model weights are distributed under a Modified MIT license, while the standalone inference library uses Apache 2.0.

These terms are not identical.

The modified model license contains conditions relevant to commercial and downstream use.

Organizations should review the complete license before redistribution, modification or integration into commercial products.

Downloadable weights do not automatically imply unrestricted use.

··········

DEVELOPERS CAN RUN TIMESTRAL WITH PYTORCH AND NUMPY.

Mistral provides a Python inference package built around PyTorch, NumPy, safetensors and Hugging Face Hub.

A forecasting pipeline typically loads the model, prepares a numerical array of historical observations and requests a prediction horizon.

The model returns forecast quantiles that can be used to construct median predictions and uncertainty intervals.

Production systems must also address timestamp alignment, sampling frequency, missing data, known-future covariates and evaluation.

A synthetic example can demonstrate the interface, but it does not validate the model on real business data.

··········

HOW TO EVALUATE TIMESTRAL AGAINST EXISTING FORECASTING SYSTEMS.

An evaluation should define the business horizon, use chronological backtesting and compare Timestral with appropriate baselines.

Point accuracy, probabilistic calibration and cost all matter.

........

Evaluation criterion	Measurement
Point accuracy	MASE, MAE or RMSE
Probabilistic quality	CRPS and interval coverage
Forecast horizon	Error by prediction distance
Robustness	Performance during structural changes
Covariate value	With-versus-without comparison
Throughput	Series per second
Latency	End-to-end processing time
Infrastructure cost	Cost per forecast
Business value	Economic impact of errors

........

A model should be selected for its contribution to actual decisions, not merely leaderboard rank.

··········

TIMESTRAL BRINGS FOUNDATION-MODEL FORECASTING TO CPU-FRIENDLY ENTERPRISE WORKLOADS.

Timestral introduces four forecasting checkpoints from 5M to 209M parameters, supporting 32,768 historical timesteps, multivariate inputs, known-future covariates and probabilistic forecasts across 21 quantiles.

Mistral reports strong results on GIFT-Eval, TIME and fev-bench, with the largest model leading its family in aggregate accuracy and smaller models offering lower latency.

The ability to run inference on ordinary server CPUs could be valuable for organizations processing large numbers of time series.

However, structural changes, forecast horizons and uncertainty calibration remain important limitations.

Timestral's practical contribution is making probabilistic foundation-model forecasting available through compact checkpoints and CPU inference, while letting organizations choose the accuracy-latency balance appropriate to their workloads.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org