Mistral released Timestral, a four-model time-series foundation model family whose 209M and 123M variants take the top two places on GIFT-Eval, TIME and fev-bench
AspireAI community · Mistral AI
Mistral announced Timestral on October 27, 2026, its first family of time-series foundation models, in four sizes: 5M, 28M, 123M and 209M parameters. The models deliver probabilistic multivariate forecasts and support both past-only and known-future covariates, and among pre-trained zero-shot models Timestral ranks first on all five metrics across the three main public forecasting benchmarks GIFT-Eval, TIME and fev-bench, with Timestral-209M and Timestral-123M taking the top two places on every metric. Notably the 28M model beats Chronos-2, a model about four times larger, on GIFT-Eval and TIME, and inference needs no GPU: forecasting one day ahead from three weeks of 15-minute history takes 17ms with the 5M model and 113ms with the 209M model on an eight-thread server CPU. The accompanying arXiv technical report (2610.37585) credits three design choices: Decoupled Horizon Forecasting that separates prediction tokens from the causally encoded context, a covariate framework supporting known-future covariates inside a causal architecture, and a two-stage pipeline that pretrains purely on a large synthetic corpus before post-training on a real-and-synthetic mix. The caveat worth stating plainly is that Mistral's leaderboard numbers are its own and the results were still pending submission to the GIFT-Eval and TIME leaderboards at announcement.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.