Introducing Timestral

Mistral announced Timestral, its first family of time-series foundation models, on its blog. The largest model, at 209 million parameters, ranks first on all three of the field's main public benchmarks — GIFT-Eval, TIME and fev-bench — according to the company, outperforming the previous state of the art while using only two-thirds of that model's parameters. Results have been submitted to the GIFT-Eval and TIME leaderboards. Mistral says the rankings cover pre-trained, zero-shot models across all five metrics and come from a single checkpoint with no ensembling.
The family ships in four sizes: 5M, 28M, 123M and 209M parameters. All four share the same architecture and training recipe, with larger models trading speed for accuracy. Mistral describes every size as sitting on the Pareto front of accuracy versus model size, meaning that for each size no public model scores better. As two concrete comparisons, the 28M model beats Chronos-2, roughly four times larger, on GIFT-Eval and TIME, while the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.
Inference does not require a GPU. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model, per Mistral. The company reports that the 28M model is five times faster than the 209M model, and that the 123M model comes within 0.2% of the 209M score at three quarters of its latency.
The models produce probabilistic, multivariate forecasts and support past-only and known-future covariates. They have a context length of 32,768 time steps, emit 21 quantiles, and carry a Modified MIT licence. Deployment is on-prem, in a VPC or in air-gapped environments. A GitHub repository is available now, with a full technical report to follow.
The technical story rests on several combined ideas. Decoupled Horizon Forecasting (DHF) departs from causal decoders trained with contiguous patch masking, as used in Toto 2.0 and TimesFM-3. In those systems the masked forecast values also drop out of the context available to later forecasts. DHF instead keeps the full history in the sequence and lets learned forecast tokens read the history before them, so every observed value acts as both target and context, real gaps stay flagged as missing, and training horizon length and missing-data augmentation can be tuned separately. Covariate-aware attention assigns each input series a role — target variate, past-only covariate or known-future covariate — with known-future covariates visible across the whole window while targets stay causal, and information flowing only from covariates to targets. That design lets a model use a holiday calendar or weather forecast without leaking a target's own future into its prediction.
Training is synthetic-first: most of the data was generated, starting with univariate processes carrying trends, seasonality and varied noise, then a multivariate synthesizer to introduce nonlinear dependencies between them. Mistral pre-trained at scale on purely synthetic data with context size 8192, then post-trained on a mix of real and synthetic data with four times the context. The objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS.
The company frames forecast types in familiar enterprise terms: sensor telemetry, observability metrics, capacity planning, energy, supply chains, risk and maintenance. It contrasts classical approaches such as ARIMA and linear regression, which are fitted to a single series and learn only from that history, and deep learning models such as PatchTST, which learn nonlinear patterns across related series but are typically trained from scratch per dataset so gains do not transfer. Time-series foundation models bring pre-training to the domain so patterns transfer zero-shot, including where data is scarce. Mistral's industrial example covers joint modelling of related sensors, quantile-based uncertainty for alert thresholds, categorical covariates such as machine state or day of week taken directly rather than encoded as ordered numbers, and known-future covariates such as production schedules.
The models are positioned to run inside existing monitoring pipelines or alongside Mistral's language models in agentic workflows — for instance flagging an anomaly, retrieving similar historical failures via embeddings, running root cause analysis with maintenance records and manuals as context, and drafting a diagnosis with evidence and next steps.
For adaptation, Timestral is available through Mistral Forge, the training pipeline used for Mistral's language models. It supports zero-shot inference, architectural extensions such as domain-specific covariates, fine-tuning on proprietary data, and control over the training data including adjusting the mix, replaying pre-training data and generating synthetic samples for the target domain. The pipeline can run on customer hardware, air-gapped included.
Why it matters: teams that forecast telemetry, demand, flows or maintenance currently either fit classical per-series models or train deep models per dataset. A 209M-parameter model that is competitive zero-shot and runs on CPU in tens of milliseconds lowers the barrier to trying foundation-model forecasting inside existing pipelines, without GPU capacity. The caveats are worth tracking: the benchmark claims are Mistral's own submissions, results are concentrated on three public leaderboards, and the technical report is still pending. In practice, the decision for most product teams is less about peak accuracy than about whether fine-tuning on in-domain data via Forge beats a well-tuned classical baseline once adaptation and operating costs are counted — a comparison the public benchmarks do not settle.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
State-of-the-art probabilistic forecasting in four sizes (5M–209M params), CPU-efficient, zero-shot ready.