Mistral's Timestral Forecasts Time Series on a Laptop CPU 🔗
Four model sizes from 5M to 209M parameters claim top benchmark rankings, with probabilistic forecasts over multivariate data and covariates.
Timestral is Mistral AI's family of time-series foundation models, built for multivariate probabilistic forecasting. Its standalone inference repository lets developers run predictions without training anything, using only PyTorch, NumPy, safetensors and huggingface_hub.
The approach is what makes it worth a closer look. Rather than fitting a separate model per dataset, a single pretrained model forecasts zero-shot, the way language models answer prompts without task-specific training. Key capabilities include:
- Joint forecasting: the variates of a series are predicted together, not one at a time.
- Covariates: past-only, known-future and categorical inputs are all supported.
- Probabilistic output: 21 quantiles per time step, so teams get uncertainty ranges, not just point estimates.
- Long context: up to 32,768 steps of history.
- CPU-friendly inference: no GPU is required, lowering the barrier for small teams.
Long horizons are handled block by block, with each block's median fed back as context. Missing values are marked as NaN or through an observation mask.
According to the launch materials, the 209M-parameter model ranked first among pretrained and zero-shot models on GIFT-Eval, TIME and fev-bench. The four sizes also sit on the accuracy-versus-size Pareto front, which means no smaller model in the family is beaten on accuracy at its footprint. Those benchmark claims come from the vendor and have not been independently reproduced here.
The code is Apache 2.0, while the weights are released separately under a Modified MIT license, so read both before commercial use.
The catch: The repository is brand new, with five open issues and no tagged release, and the benchmark rankings are self-reported at launch. Teams should validate accuracy on their own data and check how the weights' license applies to their use case before building on it.
- Demand planners forecasting inventory across many product lines at once
- Energy analysts producing probabilistic load forecasts without GPU infrastructure
- Data engineers adding zero-shot forecasting to existing monitoring pipelines
Source: mistralai/timestral — based on the project README.