Mistral Releases Timestral Time-Series Foundation Models, Smallest at 5M Parameters
jacek2023 · reddit · 2026-10-15
Mistral launched Timestral, a family of time-series foundation models for zero-shot forecasting, released in 5M, 28M, 123M, and 209M parameter versions, with source code and benchmark reproduction scripts on GitHub.
- Architecture: patch-based transformer with time attention and variate attention, 24 layers, embedding dimension 768; each patch covers 32 time steps and embeds values together with missing-value indicators.
- Capabilities: univariate and multivariate forecasting, past-only and known-future covariates (numeric or categorical, with target encoding for categoricals), missing values marked as NaN.
- Context and horizon: up to 32,768 historical time steps (1,024 patches); up to 1,024 steps (32 patches) forecast in a single forward pass, with longer horizons forecast block by block by appending each block's median forecast to the context.
- Outputs: 21 quantiles (0.01 to 0.99) for probabilistic predictions.
- Inference: PyTorch on CPU or GPU.
More from Models
- TabPFN sweeps Kaggle: now in 17% of public notebooks and 3 of September's top-10 solutions — ph_singer · 2026-10-16
- Gemini 4 Argon lands in Perplexity, claims new SoTA on WANDR benchmark — AravSrinivas · 2026-10-16
- DL Weekly #477: Haiku 5.5, building a coding agent from scratch, self-compacting agents paper — dl_weekly · 2026-10-16
- Dev says today's DLP tools are all regex-based, is rebuilding with LLM NER — HankYeomans · 2026-10-16
- Rumor: GLM 5.5 and Kimi K3.1 dropping soon, possibly slowed by anti-distillation blocks — zephyr_z9 · 2026-10-16
- Thrive Holdings benchmarks AI on real tax and IT work across 75+ firms — BorisMPower · 2026-10-16