Mistral open-sources Timestral, time-series foundation models up to 209M params
LatentSpacer · reddit · 2026-10-15
Mistral released Timestral, a family of time-series foundation models for zero-shot forecasting in 5M / 28M / 123M / 209M parameter versions.
Key details:
- Patch-based transformer with time attention and variate attention, 24 layers, embedding dim 768, 32 time steps per patch
- Univariate and multivariate forecasting, past-only and known-future covariates, probabilistic outputs across 21 quantiles (0.01–0.99)
- Context up to 32,768 time steps (1,024 patches); up to 1,024-step forecasts in a single forward pass, with longer horizons handled block by block
- Missing values marked as NaN, categorical covariates target-encoded; PyTorch inference on CPU or GPU
Weights are on Hugging Face, and Mistral open-sourced the GitHub repo to reproduce benchmarks.
More from Models
- Beam's Ten-Benchmark Breakdown: Doubles Peers on CritPt, Near-Bottom on Terminal-Bench — ArtificialAnlys · 2026-10-16
- Beam Scores 28 on Coding Agent Index, 40 Behind Leaders but Among Most Token-Frugal — ArtificialAnlys · 2026-10-16
- Beam Token Breakdown: Only 26% of Its 26k Output Tokens per Task Are Answers — ArtificialAnlys · 2026-10-16
- Reflection's First Model Beam Scores 30, Ranks #2 Among US Open Weights Models — ArtificialAnlys · 2026-10-16
- Why are sub-15B open models so rare now? Labs seem to have settled on 24-27B minimum — Hot_Example_4456 · 2026-10-16
- User spots model openly generating "ideal responses" to train other models — liminal_bardo · 2026-10-16