Mistral launches Timestral, a 209M-parameter time-series model that tops three benchmarks

Mistral AI has released Timestral, its first family of time-series foundation models. The company says the largest version, with 209M parameters, ranks first on the three main public forecasting benchmarks: GIFT-Eval, TIME and fev-bench.

What Timestral is

Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters. All share one architecture and training recipe. The models produce probabilistic forecasts as 21 quantiles, handle several related series at once, and accept two kinds of covariates: past-only ones and known-future ones such as production schedules, holidays or weather forecasts. Categorical inputs, such as a machine’s state, are taken directly rather than encoded as numbers. The context length is 32,768 time steps.

Mistral says the results come from a single checkpoint without ensembling, and that it has submitted them to the GIFT-Eval and TIME leaderboards. On GIFT-Eval it counts 83 comparable pretrained and zero-shot entries, on TIME 35 entries, and on fev-bench 37. The company also claims every size sits on the Pareto front of accuracy against model size. Its 28M model, it says, beats Chronos-2, a model about four times larger, on GIFT-Eval and TIME, and the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.

Inference does not require a GPU. Mistral measured a one-day-ahead forecast built from three weeks of 15-minute data on an 8-thread server CPU: about 10 ms for the smallest model and 140 ms for the largest. The company adds that the 123M model lands within 0.2% of the largest model’s score at three quarters of its latency.

How it was built

Mistral lists four technical ideas, with a full technical report still to come. Decoupled Horizon Forecasting keeps the whole history in the sequence and uses learned forecast tokens that read it, so every observed value serves as both target and context. Covariate-aware attention assigns each series a role, so information flows from covariates to targets without leaking a target’s own future. Most of the training data was synthetic. Training has two stages: pre-training on purely synthetic data with an 8,192 context, then post-training on a mix of real and synthetic data with a context four times longer.

Availability and customization

Mistral has published a GitHub repository for running the models on user data. The licence is listed as modified MIT, and deployment options include on-premises, VPC and air-gapped setups. Organizations that want to adapt the model can use Mistral Forge, the training pipeline the company uses for its language models. It supports architectural extensions such as domain-specific covariates, fine-tuning on proprietary data and control over the training mix, including synthetic samples. The company notes that this control lets customers make sure nothing they backtest appeared in pre-training.

Mistral also sketches a pairing with its language models: Timestral monitors plant sensor streams and flags anomalies, then an agent searches past failures and drafts a diagnosis from maintenance records and manuals.

Why it matters

The open question is how far the headline rankings carry over to real deployments, and the evidence to answer it is not yet public. Mistral has not released its technical report, and the leaderboard entries it says it submitted have not been independently confirmed. Some of the benchmarks are also narrower than a first-place claim implies. The GIFT-Eval paper describes 23 datasets with more than 144,000 series and 177 million data points. The fev-bench paper describes 100 tasks across seven domains, 46 of them with covariates. A model can lead on such aggregates and still trail on one plant’s sensor feed.

The covariate result needs the most scrutiny, since known-future inputs and categorical machine states are central to Mistral’s industrial pitch. Fewer than half of fev-bench’s tasks include covariates, so the aggregate rank says little about how well Timestral uses them. The 32,768-step context raises a similar question. Mistral’s description of the benchmark runs does not show whether long histories help on these tasks or only on the customer data it hopes to attract.

Access is the other gap. The licence is described only as modified MIT, and its terms have not been detailed in the coverage reviewed here. The Forge customization route is aimed at organizations that work with Mistral directly. Smaller teams get the repository and a model whose edge over Chronos-2 and TimesFM 2.5 they would have to test on their own data.

What’s next

Mistral says a technical report with the full results will follow. It also invites organizations with specialized forecasting problems, and researchers who want to collaborate, to get in touch.