Mistral AI announced on Thursday, October 15, 2026, the launch of Timestral, its first family of foundation models for time-series forecasting.
The largest model, with just 209 million parameters, leads the main public benchmarks in the field — GIFT-Eval, TIME and fev-bench — outperforming larger alternatives with about two-thirds of the parameters, according to the company.
What happened?
Timestral arrives in four sizes: 5M, 28M, 123M and 209M parameters. All share the same architecture and were optimized for the balance between accuracy and efficiency. The models deliver probabilistic multivariate forecasts and support past and known-future covariates.
A standout feature is the ability to run on a standard CPU. On a typical server with 8 threads, the 5M model forecasts one day ahead from three weeks of 15-minute history in about 10 ms; the 209M model takes about 140 ms.
Why it matters
Time series are everywhere in the real world: industrial sensors, observability metrics, capacity planning, energy, supply chains and finance. Classical models like ARIMA or networks trained from scratch on each dataset do not transfer learning well across domains. Time-series foundation models like Timestral learn generic patterns in pre-training and perform zero-shot forecasting.
Mistral highlights uses such as anomaly detection in industrial plants, where the model can jointly forecast vibration, temperature and pressure, providing uncertainty quantiles instead of a single number.
In practice
Weights are available on Hugging Face and code on GitHub under Modified MIT for weights and Apache 2.0 for the code. The company offers fine-tuning via Mistral Forge for adaptation to proprietary data, including in air-gapped environments.
A full technical report is expected soon.
Sources
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot
