Drafted with AI tools and edited by a human. The figures and conclusions were checked by a GPUniq editor.
TL;DR
Mistral AI released Timestral on October 15, 2026. The family provides zero-shot probabilistic forecasting across four sizes scaling up to 209 million parameters. It handles up to 32,768 historical timesteps. The flagship Timestral-209M scores 0.658 on GIFT-Eval MASE and 0.522 on TIME CRPS.
What is the API pricing?
Mistral AI charges credits per forecast call. The web demo is free with quota limits. API forecast calls use credits. You should check the vendor dashboard for current credit rates.
What model sizes are available?
The family spans four parameter tiers to fit different hardware footprints.
| Model Tier | Parameter Scale | Target Deployment | Inference Hardware |
|---|---|---|---|
| Timestral-5M | 5 Million | - | - |
| Timestral-28M | 28 Million | - | - |
| Timestral-123M | 123 Million | - | - |
| Timestral-209M | 209 Million | - | - |
Smaller variants like Timestral-5M run efficiently on standard CPU hardware.
How does Timestral perform on benchmarks?
Data from the mistralai/Timestral-209M Hugging Face repository shows evaluation scores across three metrics. Lower scores mean better error rates. Higher scores show better task skill.
| Benchmark Name | Metric Type | Score |
|---|---|---|
| GIFT-Eval MASE | Lower is better | 0.658 |
| TIME CRPS | Lower is better | 0.522 |
| fev-bench SQL skill score | Higher is better | 50.39 |
What is the maximum context?
Models process up to 32,768 historical timesteps in a single pass. This long-range context handles complex temporal dependencies.
How do I generate quantile outputs?
Each timestep yields 21 distinct quantile outputs to map out prediction intervals. Probabilistic modeling works out-of-the-box by parsing these quantile tensors.
How do I deploy Timestral?
You can download open source weights from the mistralai/Timestral-209M Hugging Face repository. Local inference runs using standard PyTorch pipelines on CPU or GPU hardware. Teams can also deploy the models on-premises and in air-gapped environments.
What are the limitations?
The architecture has specific constraints you need to plan for. It does not foresee novel events without precedent in training data. Smaller models may lose seasonal patterns when data is very noisy. It struggles on cleanly growing counters versus simpler baselines. Finally, generating a forecast horizon beyond one decode block requires block-wise forecasting.
Related on the GPUniq blog
Methodology
Facts are compiled from the vendor's own announcement and documentation, linked below, at the time of writing; where the vendor hasn't published a number it is marked as such. GPUniq prices, when the model is already available here, come from our live catalog.
Sources
- 1.Timestral: Mistral’s time-series foundation models — Mistral AI (accessed )
- 2.Timestral 1.0 - Mistral Docs — Mistral AI (accessed )
Want cheap GPUs for your next project?
Browse live GPU prices and rent the right card in seconds — H100, A100, RTX 4090, and 50+ more models.
