Mistral AI Releases Timestral: A 209M Forecasting Model Family That Outperforms Google’s TimesFM-3

Why it matters

Mistral AI has launched Timestral, its first family of time series foundation models, offering four dense checkpoints that forecast unseen data zero-shot after pretraining on a large, mostly synthetic corpus. The largest checkpoint ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench, according to Mistral's own reported scores, edging out Google's TimesFM-3 and t0-gamma. A key practical point is that inference runs on standard server CPUs without a GPU, with latency scaling across the model sizes. Architecturally, Timestral is a causal, patch-based transformer using decoupled horizon forecasting, covariate-aware attention to prevent future leakage, and two-stage training with pinball loss over quantiles. Unlike classical tools such as ARIMA or deep models like PatchTST that fit or train per dataset, Timestral transfers patterns across domains. Weights are open under a modified MIT licence, contrasting with TimesFM-3's non-commercial terms, and fine-tuning is available through Mistral Forge with deployment options including on-prem and air-gapped environments.