Mistral’s Timestral Brings Fast, GPU-Free Forecasting to Four Model Sizes
Mistral AI released Timestral on October 18, 2026, introducing a family of four zero-shot forecasting models built for time series data. The models range from 5 million to 209 million parameters, giving teams several options for balancing forecast quality, download size, and response time.
The largest model, Timestral-209M, ranks first among pretrained, zero-shot models on GIFT-Eval, TIME, and fev-bench. It also produces forecasts without a GPU at inference, which makes the family suited to setups where predictions must run on standard server hardware.
Four models, one forecasting system
Timestral comes in four sizes: 5M, 28M, 123M, and 209M parameters. The 209M version contains 837 MB of weights in fp32, while the listed download sizes are 21 MB for the 5M model, 111 MB for the 28M model, and 490 MB for the 123M model.
That range matters because forecasting workloads do not all have the same needs. A smaller model can reduce download and response costs, while a larger model can target benchmark performance. Timestral’s size scaling is flat at the top, and the 123M model has latency within 0.2% of the 209M model’s score.
The 28M model also beats Chronos-2 on GIFT-Eval and TIME. Chronos-2 comes from Amazon, while Timestral comes from Mistral AI. The comparison gives the smaller Timestral checkpoint a clear role for users who want stronger benchmark results without moving to the largest model.
Forecasts designed for real-world time series
Timestral produces probabilistic forecasts rather than a single number for each future point. It returns 21 quantiles per time step, giving users a range of possible outcomes instead of one fixed prediction.
The models forecast multivariate targets together and accept three types of covariates: past-only data, known-future data, and categorical data. That lets Timestral use more than the target’s history when those extra signals are available.
Technically, Timestral is a causal, patch-based transformer with time attention and variate attention. Each patch covers 32 time steps, and one forward pass can forecast up to 1,024 steps. These design choices let the model process sections of a series while still tracking relationships across time and variables.
Benchmark scores and CPU speed
Timestral-209M posted a MASE of 0.658 on GIFT-Eval, with a CRPS of 0.446. On TIME, it recorded a MASE of 0.626 and a CRPS of 0.522. On fev-bench, it reached a SQL skill score of 50.39 and a MASE skill score of 39.49.
Those results support Mistral AI’s claim that the largest checkpoint ranks first among pretrained, zero-shot models across the three benchmark tests. The 28M model’s results against Chronos-2 also show that benchmark performance does not depend only on choosing the biggest available checkpoint.
Speed is another part of the Timestral pitch. Timestral-209M took 140 milliseconds on a standard server CPU with 8 threads to forecast 1 day ahead from 3 weeks of 15-minute history. The 5M model completed the same task in 10 milliseconds, while the 28M model is 5 times faster than the 209M model.
That performance does not require a GPU during inference. For forecasting systems that handle many repeated requests, the difference between 10 milliseconds, the speed of the 28M model, and 140 milliseconds can shape which model fits the workload.
What sits behind Timestral
Mistral AI lists four technical contributions behind the model family: Decoupled Horizon Forecasting, Covariate-aware attention, Synthetic-first data, and Two-stage training. Together, these contributions describe how Timestral handles forecast horizons, additional data inputs, training data, and the training process.
The 209M model uses 24 layers and an embedding dimension of 768. Its patch structure covers 32 time steps at a time, while the forecasting system supports up to 1,024 future steps in a single forward pass.
For users choosing between the four checkpoints, the numbers point to different tradeoffs. The 5M model offers the smallest download and the 10-millisecond result, the 28M model combines a lower size with a stated advantage over Chronos-2 on GIFT-Eval and TIME, and the 123M model stays close to the 209M model’s latency score. The 209M model remains the benchmark leader among the Timestral options listed here.
Timestral’s release puts zero-shot time series forecasting into a compact family rather than a single fixed model. With probabilistic outputs, support for past and future covariates, multivariate forecasting, and CPU inference, the four checkpoints target different ways of putting forecasts into use.
