SHELFCRITTER AI Digest

MODEL

Timestral

modeltopic-notemistral

Overview

Timestral is Mistral‘s open-weight time-series foundation-model family released as arXiv:2610.37585, with four variants at 5M / 28M / 123M / 209M parameters and weights published on Hugging Face. The models are trained via a two-stage synthetic-then-real pipeline. The two largest variants take the top two places on every metric across the GIFT-Eval, TIME, and fev-bench time-series forecasting benchmarks — Mistral staking a claim in a vertical where open-weight FMs have been sparse.

Timeline

  • 2026-10-29-AI-Digest — Mistral releases Timestral: four open-weight time-series foundation models (5M / 28M / 123M / 209M params) trained via a two-stage synthetic-then-real pipeline, with weights on Hugging Face (arXiv:2610.37585). The two largest variants take the top two places on every metric across GIFT-Eval, TIME, and fev-bench. Why it matters: practitioner-relevant for forecasting workloads, and extends Mistral’s differentiation-outside-the-mainstream-benchmark-race pattern (already visible in Leanstral 1.5 theorem-proving, Robostral Navigate single-camera navigation, and Shieldstral multimodal moderation) with a time-series foundation-model lane.

Key Developments

  1. Open-Weight Time-Series FM Family at Four Sizes (5M / 28M / 123M / 209M), Top of GIFT-Eval / TIME / fev-bench (October 29, 2026): Mistral ships Timestral as four open-weight checkpoints trained via a two-stage synthetic-then-real pipeline, with the two largest variants topping every metric across GIFT-Eval, TIME, and fev-bench. Weights published on Hugging Face per the arXiv paper (arXiv:2610.37585). Load-bearing framing to carry: open-weight time-series FMs have been sparse — Timestral is Mistral staking a claim in a vertical that practitioners doing forecasting workloads could not previously reach for an open-weight substrate on.

  2. Fifth Off-Mainstream Lane in Mistral’s Differentiation Strategy (October 29, 2026): Timestral extends the pattern the corpus has been tracking across Leanstral 1.5 (formal-math theorem-proving), Robostral Navigate (single-camera robotic navigation), Shieldstral (multimodal moderation), and Mistral Large 4 (frontier-tier) into a time-series FM lane. Reframe worth carrying: Mistral’s differentiation keeps compounding across off-mainstream lanes rather than contesting the closed-frontier reasoning / SWE-bench cohort head-on — a fifth disciplinary bet rather than head-on frontier competition.

See also: Mistral, Leanstral 1.5, Shieldstral, Mistral Large 4, MOC - Open Source Models.