Mistral released Timestral on October 15, 2026, its first family of time-series foundation models. The company says the largest model, with 209M parameters, ranks first among pre-trained, zero-shot models on all five metrics across GIFT-Eval, TIME and fev-bench, outperforming the previous state of the art with two-thirds as many parameters. Results were submitted to the GIFT-Eval and TIME leaderboards. Timestral produces probabilistic, multivariate forecasts and accepts past-only and known-future covariates (extra variables used to inform forecasts). Mistral says all four model sizes sit on the accuracy-versus-size Pareto front, meaning no public model forecasts better at each size.

Time-series data includes enterprise sensor readings and monitoring metrics used in capacity planning, energy, supply chains, risk and maintenance. Unlike ARIMA and linear regression, which are generally fitted to one series and learn from its history, deep-learning models such as PatchTST can learn nonlinear patterns across related series but are typically trained from scratch for each dataset. Mistral says Timestral is pre-trained on a broad corpus to transfer learned patterns across series and make zero-shot forecasts on unseen data, including in data-scarce domains. Its benchmark results use one model checkpoint without combining multiple models.

The Timestral family has four sizes: 5M, 28M, 123M and 209M parameters. Larger models improve forecast performance; smaller ones are faster and cheaper. Mistral says its 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME, while the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. On a standard server CPU using 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms for the 5M model and 140 ms for the 209M model. The 28M model is five times faster than the 209M model; the 123M model scores within 0.2% of the 209M model at three quarters of its latency.

For an industrial plant, Timestral can forecast related sensor series jointly, such as a pump’s vibration, temperature and pressure. It predicts quantiles to show forecast uncertainty, accepts category inputs such as weekday or machine state, and can use known-future information such as production schedules, planned shutdowns and weather forecasts. Mistral describes a workflow in which Timestral monitors sensor streams for anomalies, then an agent compares signal embeddings with historical sensor data. Sub-agents could analyze maintenance records, shift logs, operator notes, manuals, work orders and related equipment readings; the agent then drafts a diagnosis with evidence and possible next steps.

Mistral says Timestral can run in existing monitoring pipelines or work with its language models in multimodal agentic workflows. Its technical report will detail four methods. Decoupled Horizon Forecasting keeps the full history available while learned forecast tokens read it, so observed values serve as both targets and context; true gaps remain marked as missing. Covariate-aware attention assigns each input a role and lets known-future covariates inform targets without exposing a target’s future values. Most training data was synthetic, generated first as univariate processes with trends, seasonality and varied noise, then combined into multivariate series with nonlinear dependencies. Training begins on synthetic data with context size 8192, then continues on real and synthetic data with a 4x longer context; the objective is pinball loss over 21 quantiles, a discrete approximation of CRPS.

Organizations can use Mistral Forge for zero-shot inference, architectural extensions such as domain-specific covariates, fine-tuning on proprietary data, and control over training data, including its mix, replayed pre-training data and generated synthetic samples. Mistral says the pipeline can run on client hardware, including air-gapped environments. The company published a GitHub repository to run Timestral on users’ data and said a technical report with full results will follow. The model has a context length of 32,768 time steps, outputs probabilistic forecasts with 21 quantiles, accepts multivariate targets and past-only or known-future covariates, and can be deployed on-premises, in a VPC or in an air-gapped environment under a modified MIT license.

Mistral released Timestral on October 15, 2026, its first family of time-series foundation models. The company says the largest model, with 209M parameters, ranks first among pre-trained, zero-shot models on all five metrics across GIFT-Eval, TIME and fev-bench, outperforming the previous state of the art with two-thirds as many parameters. Results were submitted to the GIFT-Eval and TIME leaderboards. Timestral produces probabilistic, multivariate forecasts and accepts past-only and known-future covariates (extra variables used to inform forecasts). Mistral says all four model sizes sit on the accuracy-versus-size Pareto front, meaning no public model forecasts better at each size.

Time-series data includes enterprise sensor readings and monitoring metrics used in capacity planning, energy, supply chains, risk and maintenance. Unlike ARIMA and linear regression, which are generally fitted to one series and learn from its history, deep-learning models such as PatchTST can learn nonlinear patterns across related series but are typically trained from scratch for each dataset. Mistral says Timestral is pre-trained on a broad corpus to transfer learned patterns across series and make zero-shot forecasts on unseen data, including in data-scarce domains. Its benchmark results use one model checkpoint without combining multiple models.

The Timestral family has four sizes: 5M, 28M, 123M and 209M parameters. Larger models improve forecast performance; smaller ones are faster and cheaper. Mistral says its 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME, while the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. On a standard server CPU using 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms for the 5M model and 140 ms for the 209M model. The 28M model is five times faster than the 209M model; the 123M model scores within 0.2% of the 209M model at three quarters of its latency.

For an industrial plant, Timestral can forecast related sensor series jointly, such as a pump’s vibration, temperature and pressure. It predicts quantiles to show forecast uncertainty, accepts category inputs such as weekday or machine state, and can use known-future information such as production schedules, planned shutdowns and weather forecasts. Mistral describes a workflow in which Timestral monitors sensor streams for anomalies, then an agent compares signal embeddings with historical sensor data. Sub-agents could analyze maintenance records, shift logs, operator notes, manuals, work orders and related equipment readings; the agent then drafts a diagnosis with evidence and possible next steps.

Mistral says Timestral can run in existing monitoring pipelines or work with its language models in multimodal agentic workflows. Its technical report will detail four methods. Decoupled Horizon Forecasting keeps the full history available while learned forecast tokens read it, so observed values serve as both targets and context; true gaps remain marked as missing. Covariate-aware attention assigns each input a role and lets known-future covariates inform targets without exposing a target’s future values. Most training data was synthetic, generated first as univariate processes with trends, seasonality and varied noise, then combined into multivariate series with nonlinear dependencies. Training begins on synthetic data with context size 8192, then continues on real and synthetic data with a 4x longer context; the objective is pinball loss over 21 quantiles, a discrete approximation of CRPS.

Organizations can use Mistral Forge for zero-shot inference, architectural extensions such as domain-specific covariates, fine-tuning on proprietary data, and control over training data, including its mix, replayed pre-training data and generated synthetic samples. Mistral says the pipeline can run on client hardware, including air-gapped environments. The company published a GitHub repository to run Timestral on users’ data and said a technical report with full results will follow. The model has a context length of 32,768 time steps, outputs probabilistic forecasts with 21 quantiles, accepts multivariate targets and past-only or known-future covariates, and can be deployed on-premises, in a VPC or in an air-gapped environment under a modified MIT license.