Introducing Timestral
Mistral
Mistral just released Timestral, a new model family for forecasting time-series data. It runs on CPUs, beats larger rivals on key benchmarks, and can be used on-prem or even air-gapped.
Based on reporting by Mistral — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral has released Timestral, its first family of time-series foundation models, and the pitch is simple: smaller models, strong forecasts, less GPU fuss. The largest version has 209 million parameters, and Mistral says it ranks first on the field’s three main public benchmarks: GIFT-Eval, TIME and fev-bench. It also says the model does that with about two-thirds the parameters of the previous state of the art.
That matters because time-series forecasting is everywhere in enterprise work, from sensor telemetry to observability metrics. The uses Mistral names are the unglamorous but expensive ones: capacity planning, energy, supply chains, risk and maintenance. Timestral is meant to fit that world. It produces probabilistic, multivariate forecasts, handles past-only and known-future covariates, and is designed to work without a GPU at inference.
The family comes in four sizes: 5M, 28M, 123M and 209M parameters. Mistral says the smaller models are faster and cheaper, while the larger ones improve accuracy, and that each size sits on the Pareto front for forecast quality versus model size. On a standard server CPU with 8 threads, the company says a one-day-ahead forecast from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model. It also says the 123M version gets within 0.2% of the 209M score at three quarters of its latency.
Under the hood, Timestral leans heavily on synthetic data and a two-stage training recipe. Mistral says it pre-trained first on purely synthetic sequences with a context size of 8192, then post-trained on a mix of real and synthetic data with four times longer context. The model uses decoupled horizon forecasting and covariate-aware attention so future-known signals like schedules or weather forecasts can help without leaking target values into the prediction.
The other notable piece is deployment. Timestral is available through Mistral Forge for zero-shot inference, fine-tuning, domain-specific covariates and data control, including replaying pre-training data and generating synthetic samples. Mistral says the full pipeline can run on client hardware, including air-gapped environments. For a company selling forecasting models, that’s probably the real hook: not just accuracy, but a way to keep the thing inside the machine room.
My take — AI-written commentary, not fact-checked reporting
The interesting part here isn’t that Mistral built a forecasting model; it’s that they built one that seems aimed at boring, expensive infrastructure instead of demo theater. CPU inference, on-prem deployment and air-gapped support are the kind of details that usually matter more than benchmark trophies once the slide deck closes. Also, synthetic-first training sounds less sexy than “magic,” which is often a good sign.
Read more about this at: Mistral