Mistral AI Launches Timestral, a Family of Time-Series Forecasting Models
Mistral AI released Timestral, a four-size family of time-series foundation models it says tops three public forecasting benchmarks.
Quick answer
What did Mistral AI announce with Timestral?
Mistral AI released Timestral, its first family of time-series foundation models, in four sizes from 5 million to 209 million parameters. The company says its largest model ranks first on three public forecasting benchmarks and that the models can run on CPUs without requiring a GPU.
Key takeaways
- Mistral AI released Timestral, its first family of time-series foundation models, on October 15, 2026.
- The family includes four model sizes: 5M, 28M, 123M and 209M parameters, all sharing the same architecture and training recipe.
- Mistral says its 209M-parameter model ranks first on all three main public time-series forecasting benchmarks: GIFT-Eval, TIME and fev-bench.
- The company says the models run on CPUs without needing a GPU, and support multivariate, probabilistic forecasts with past and future covariates.
- Timestral is available through Mistral Forge for fine-tuning, and a GitHub repository has been published for running the models on user data.
Mistral AI released Timestral on October 15, 2026, a new family of time-series foundation models built to generate forecasts for enterprise data such as sensor telemetry, energy usage, and financial flows.
What Mistral announced
Timestral comes in four sizes: 5 million, 28 million, 123 million and 209 million parameters. All four share the same architecture and training recipe, according to Mistral. The company says the largest model ranks first on all three of what it calls the field's main public benchmarks — GIFT-Eval, TIME and fev-bench — outperforming the previous state of the art while using about two-thirds of the parameters. Mistral says these results come from a single checkpoint, without ensembling, and that it has submitted results to the GIFT-Eval and TIME leaderboards.
Mistral says the models produce probabilistic, multivariate forecasts and can take in past-only and known-future covariates, such as production schedules or weather forecasts, to improve prediction quality. The company also says the models can handle categorical inputs, such as a machine's operating state, without converting them into numbers that would impose an artificial order.
Technical details
According to Mistral, Timestral does not require a GPU for inference. The company says that on a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 milliseconds with the 5M model and 140 milliseconds with the 209M model. Mistral says the 28M model is five times faster than the 209M model, and that the 123M model comes within 0.2% of the 209M model's benchmark score at three-quarters of its latency.
Mistral attributes the results to several technical contributions it says will be detailed in a forthcoming technical report. These include a method the company calls Decoupled Horizon Forecasting, which it says keeps the full history in the training sequence while learned forecast tokens read from that history, rather than masking out forecast values as missing data. The company also describes a covariate-aware attention mechanism that keeps known-future covariates visible across the input window while restricting information flow from covariates to targets, and says the models were trained primarily on synthetic data generated to mimic realistic temporal patterns before a second training stage mixed in real data with a longer context window.
Mistral lists the models' context length at 32,768 time steps, with outputs given as probabilistic forecasts across 21 quantiles. The company says the models are released under a modified MIT license and can be deployed on-premises, in a VPC, or in air-gapped environments.
Availability and adaptation
Mistral says Timestral is available through Mistral Forge, the same training pipeline the company uses for its language models, allowing customers to fine-tune the models on proprietary data, add domain-specific covariates, and control the training data mix. The company says this pipeline can be deployed on a client's own hardware, including air-gapped environments.
Mistral has published a GitHub repository so users can run Timestral on their own data, and the company says a technical report with full results will follow.
Source: Mistral AI, "Introducing Timestral," published October 15, 2026.
Frequently asked questions
- What is Timestral?
- Timestral is Mistral AI's first family of time-series foundation models, designed to produce probabilistic, multivariate forecasts.
- What sizes does Timestral come in?
- Mistral released four sizes: 5M, 28M, 123M and 209M parameters, all sharing the same architecture and training recipe.
- How does Mistral say Timestral performs on benchmarks?
- Mistral says its 209M-parameter model ranks first on all three main public forecasting benchmarks, GIFT-Eval, TIME and fev-bench, outperforming the prior state of the art with fewer parameters.
- Does Timestral require a GPU to run?
- No. Mistral says the models run on a standard CPU, with the 5M model forecasting in 10 milliseconds and the 209M model in 140 milliseconds for a one-day-ahead forecast in its stated test.