Mistral Open-Sources Timestral Time-Series Foundation Models Up to 209M Parameters
On October 14–15, Mistral released and open-sourced the Timestral time-series foundation model family on Hugging Face, spanning four parameter sizes (5M, 28M, 123M, 209M) and focused on zero-shot time-series forecasting. This is Mistral's first time-series forecasting foundation model on Hugging Face, with weights provided in safetensors format plus open-source code and benchmark reproduction scripts.
Confirmed
- The models were officially released by Mistral AI on Hugging Face, including Timestral-5M and Timestral-209M, using the time-series-forecasting pipeline, with weights in safetensors format under a custom license
- The architecture is a patch-based Transformer supporting zero-shot forecasting
- Supports 21 quantile outputs with a context length of up to 32k
- Achieves zero-shot SOTA on 3 major time-series forecasting benchmarks
- Code and benchmark reproduction scripts are open-sourced alongside
Why it matters
Time-series forecasting has long been dominated by statistical methods and small specialized models. Timestral brings the "foundation model + zero-shot" paradigm to the field and is open-sourced at lightweight scales from 5M to 209M parameters, lowering the compute barrier for enterprise adoption—a significant signal that a leading European model vendor is expanding beyond language models.
2026-10-15 ~ 2026-10-16 · 6 related posts
Primary sources
- [source] Mistral ships Timestral-209M, a 209M time-series forecasting foundation model — mistralai · 2026-10-14
- Mistral Releases Timestral Time-Series Foundation Models, Smallest at 5M Parameters — jacek2023 · 2026-10-15
- Mistral open-sources Timestral, time-series foundation models up to 209M params — LatentSpacer · 2026-10-15
- Mistral releases Timestral, a time series foundation model with SOTA zero-shot forecasting — _akhaliq · 2026-10-16