Skip to content
NewsModel release

Mistral ships Timestral, a forecasting model that beats bigger rivals on CPU

· by Pondero Newsdesk

The short version

Mistral's new Timestral family forecasts time-series data without a GPU, and its 209M-parameter model outperforms larger specialist forecasters on three public benchmarks.

Mistral ships Timestral, a forecasting model that beats bigger rivals on CPU

Mistral released Timestral on October 15, 2026, a family of models built specifically for time-series forecasting rather than a general chat model repurposed for the job, and said its largest variant beats the prior state of the art while using two-thirds the parameters, per Mistral's announcement.

What

Timestral ships in four sizes, 5M, 28M, 123M, and 209M parameters, all sharing the same architecture and training recipe. Mistral's largest variant ranks first on the field's three main public benchmarks (GIFT-Eval, TIME, fev-bench), according to the company. A quarter its size, the 28M model still beats the specialist forecaster Chronos-2 on GIFT-Eval and TIME. Every size in the family produces probabilistic forecasts across 21 quantiles, handles multivariate series with past-only and known-future covariates, and reads up to 32,768 time steps of context, Mistral said. None of the models need a GPU: an 8-thread server CPU returns a forecast in 10 milliseconds with the 5M model and 140 milliseconds with the 209M model, per the same post. Code is live now under a modified MIT license on GitHub, and the weights are published on Hugging Face.

CPU-only inference removes the GPU bottleneck for continuous forecasting

Teams running forecasts across thousands of sensors or accounts, the kind of workload Mistral points to in energy and manufacturing, have typically needed either a classical model fit to each series individually or a GPU cluster to run a deep model at scale. Timestral's sub-200-millisecond CPU inference lets that forecasting run on infrastructure those teams already operate, without provisioning accelerators. Fine-tuning and architectural extensions are available through Mistral Forge, with deployment on-premises or in a fully air-gapped VPC, aimed at regulated industries unwilling to send sensor or transaction data to a third-party API.

Context and reactions

Mistral was explicit that the benchmark results come from a single checkpoint with no ensembling, and that they have been submitted to the GIFT-Eval and TIME leaderboards rather than independently verified there yet. Recent forecasters like Toto 2.0 and TimesFM-3 mask out the values being predicted, which also removes them as context for later steps. Timestral keeps the full history in the sequence instead, reading it through separate forecast tokens. Training leaned heavily on synthetic data, Mistral said, generating univariate processes first and combining them into multivariate series with engineered dependencies before a second stage mixed in real data.

What to watch next

Whether the GIFT-Eval and TIME leaderboards post Timestral's submitted entries publicly matters, along with whether independent teams reproduce the gap Mistral reported over Chronos-2, TimesFM-3, and Falcon-2.0. Adoption signals, such as energy trading or manufacturing planning teams publishing results from the open GitHub release, would be the next real test.

Sources