Mistral introduces open-weight Timestral models for forecasting capacity, energy and sensor data

Mistral AI announced Timestral, a set of time-series forecasting models, on 15 October 2026. According to the company's blog post, Timestral is Mistral's first foundation model family in this field and includes four models with 5 million, 28 million, 123 million and 209 million parameters. The models produce probabilistic, multivariate forecasts and can take known future variables as input alongside past data.

• Forecasts come with 21 quantiles, and the context length is 32,768 time steps.
• Mistral says the 209-million-parameter model ranks first among pre-trained, zero-shot models on the GIFT-Eval, TIME and fev-bench benchmarks.
• In Mistral's speed test on a standard server CPU with 8 threads, predicting the next day from three weeks of readings taken every 15 minutes took 17 milliseconds on the 5M model and 113 milliseconds on the 209M model.
• All four models are listed in a Timestral collection on Hugging Face under a modified MIT license; they can be fine-tuned through Mistral Forge and deployed on-premises, in a VPC or air-gapped.

Why it matters

Mistral lists sensor telemetry, capacity planning, supply chains, energy, maintenance and finance as use cases for Timestral. These are everyday planning topics for manufacturers, energy-intensive plants and inventory-holding companies in Türkiye. Because the models are small and can forecast in milliseconds on an ordinary server without a GPU, testing them on ERP or IoT data could be relatively easy. The results are Mistral's own benchmark figures and should be checked against a company's own historical data.