Mistral releases Timestral time-series foundation models, tops GIFT-Eval
mistralai · hf · 2026-10-28
Mistral AI introduces Timestral, a family of four time-series foundation models (5M/28M/123M/209M) for probabilistic forecasting.
- Sets a new SOTA on GIFT-Eval, TIME, and fev-bench, with Timestral-209M and 123M taking the top two places on every metric
- Three key ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context so longer horizons can be trained without masking context, with each observed value serving as both forecast target and context; a covariate framework natively supporting known future covariates; synthetic pretraining followed by mixed post-training
A notable open release for teams working on time-series forecasting.
More from Models
- AI labs near IPO, yet LLMs still produce 'astonishingly terrible prose' by default — GabGarrett · 2026-10-28
- Rebuttal: Kimi 3.1 Nears Fable on Frontier Code 2, Contradicting 'China Models Crushed' Claim — bookwormengr · 2026-10-28
- Open model Flint hits #2 trending on Hugging Face, team outlines next steps — josh_tobin_ · 2026-10-28
- YacineMTB rails: even the best LLMs are 'complete dogshit' and he's tired of it — yacineMTB · 2026-10-28
- Omni-InteractScope benchmarks proactive interaction in continuous omnimodal streams — Qwen · 2026-10-28
- Gemini 4 Argon Hits 1.73% WER on OCR Benchmark, Surpassing Human Experts — gabriberton · 2026-10-28