Mistral launches Timestral time-series foundation models, zero-shot SOTA on GIFT-Eval
victormustar · x · 2026-10-20
Mistral AI released Timestral, a family of time-series foundation models for zero-shot forecasting in 5M/28M/123M/209M sizes. A dev demo lets you draw a curve and watch the model predict what comes next. Key details: patch-based transformer with time and variate attention, 32 time steps per patch, context up to 32,768 steps, forecasts up to 1,024 steps per forward pass, 21 quantile outputs, and support for covariates and NaN-marked missing values. The 209M model achieves state-of-the-art zero-shot accuracy on GIFT-Eval, TIME, and fev-bench, and every size leads its model class (GIFT-Eval MASE: 0.705 for 5M down to 0.658 for 209M).
Related event: Mistral Unveils Timestral Time-Series Foundation Models(2 posts)→
More from Models
- NVIDIA's Looped Dory1.0-4B lets you dial inference depth from 1 to 4 passes — lmoroney · 2026-10-20
- NVIDIA's Dory1.0-4B Looped Model Lets You Pick Inference Depth at Runtime — lmoroney · 2026-10-20
- OpenAI dropped 722 AI-generated math papers in a day; can humans read them? — songyoupeng · 2026-10-20
- JevBench adds 34 decision models in four days; Nimble-v4-max price cut to $0.04/1M tokens — airesearch12 · 2026-10-20
- Mercury 3 benchmarks: beats GPT-6 Luna and Haiku 5.5 at 1,068 tok/s — StefanoErmon · 2026-10-20
- Inception Labs releases Mercury 3, a diffusion LLM claiming 1,000+ tok/s at GPT-6-level quality — StefanoErmon · 2026-10-20