Timestral
Mistral's first time-series forecasting models: four CPU-friendly sizes (5M-209M) with probabilistic output and open weights.
Overview
Timestral is Mistral AI's first family of time-series foundation models, released October 15, 2026 in four sizes of 5M, 28M, 123M and 209M parameters that share one architecture. It produces probabilistic forecasts with 21 quantiles from multivariate targets plus past-only and known-future covariates, over a 32,768-step context. Mistral reports that the 209M model ranks first among pre-trained zero-shot models on GIFT-Eval, TIME and fev-bench, and that the 28M model beats Chronos-2 (about four times larger) on GIFT-Eval and TIME. Inference needs no GPU: a one-day-ahead forecast takes about 10 ms on the 5M model and 140 ms on the 209M model on an 8-thread server CPU.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Four sizes: 5M, 28M, 123M and 209M parameters
- 32,768-step context with multivariate targets and past/future covariates
- Probabilistic output with 21 quantiles
- CPU-only inference: ~10 ms (5M) to ~140 ms (209M) per one-day-ahead forecast
- Top zero-shot rank on GIFT-Eval, TIME and fev-bench for the 209M model (Mistral-reported)
- Deployable on-prem, in a VPC or air-gapped; weights on Hugging Face under a modified MIT license
- Fine-tuning and customization through Mistral Forge
Use cases
- Industrial monitoring: anomaly detection and predictive maintenance
- Finance: cash flows, volumes, volatility and treasury forecasting
- Capacity planning in energy and supply chains
Pricing
Weights downloadable under a modified MIT license; hosted pricing not stated.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 15 Oct 2026
- Record updated
- 19 Oct 2026
- Record last checked
- 19 Oct 2026