Mistral AI Releases Timestral Time-Series Models, Claiming Zero-Shot Benchmark Lead

The largest checkpoint, Timestral-209M, ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench, according to Mistral-reported results cited by MarkTechPost. On fev-bench, it records an SQL skill score of 50.39, ahead of TimesFM-3 at 48.99. Its narrowest lead in the reported figures is on GIFT-Eval, where its MASE of 0.658 edges t0-gamma at 0.661 by 0.003. On GIFT-Eval and TIME, lower MASE and CRPS are better; on fev-bench, higher skill scores are better. All scores are zero-shot from a single checkpoint without ensembling, according to the Timestral-209M model card.

Timestral is a causal, patch-based transformer with time attention and variate attention. The 209M model uses 24 layers and an embedding dimension of 768, with each patch covering 32 time steps. It accepts context up to 32,768 time steps and a single forward pass forecasts up to 1,024 steps; longer horizons run block by block, feeding each block's median back as context. Each model outputs 21 quantiles per time step, forecasts multivariate targets jointly and accepts past-only, known-future and categorical covariates. Mistral lists four technical contributions: Decoupled Horizon Forecasting, covariate-aware attention that lets information flow only from covariates to targets, synthetic-first training data and two-stage training with pinball loss over 21 quantiles.

On benchmark details listed in the model card, Timestral-209M records MASE 0.658 and CRPS 0.446 on GIFT-Eval, MASE 0.626 and CRPS 0.522 on TIME, and an SQL skill score of 50.39 with an MASE skill score of 39.49 on fev-bench. Scaling flattens at the top: Timestral-123M scores 0.659 MASE on GIFT-Eval against 0.658 for the 209M model. Mistral's team states that the 28M model beats Chronos-2 on GIFT-Eval and TIME, even though Chronos-2 is about four times larger.

On a standard server CPU with eight threads, forecasting one day ahead from three weeks of 15-minute history took 10 milliseconds with the 5M model and 140 milliseconds with the 209M model, according to Mistral's tests. The 28M model is five times faster than the 209M model, and the 123M model lands within 0.2% of the 209M score at three quarters of its latency. The 209M weights are 837 MB in fp32, and Mistral says a GPU is optional. Download sizes listed for the smaller models are 21 MB for 5M, 111 MB for 28M and 490 MB for 123M.

The model card compares Timestral-209M with TimesFM-3 at 330M parameters from Google Research, Chronos-2 at 120M from Amazon and Toto-2.0-2.5B from Datadog. It lists Timestral under a modified MIT license, TimesFM-3 under a non-commercial v1.0 license, and Chronos-2 and Toto 2.0 under Apache 2.0. Google and Amazon do not publish matching figures on their own model pages, and the card's Toto 2.0 TIME scores differ from Datadog's model card: 0.642 and 0.539 on Mistral's card versus 0.640 and 0.532 on Datadog's.

Developers can install the package with pip install git+https://github.com/mistralai/timestral.git. It requires Python 3.10 or later and torch 2.4 or later, with dependencies on PyTorch, NumPy, safetensors and huggingface_hub. Weights download from Hugging Face. Mistral's documentation lists Timestral 1.0 209M as a Public Preview, and fine-tuning on proprietary data runs through Mistral Forge. The pipeline can deploy on-premises, in a VPC or air-gapped.

Editor's Summary

Mistral AI's Timestral family spans 5M to 209M parameters and is positioned for CPU-only zero-shot forecasting, with Mistral-reported leads on GIFT-Eval, TIME and fev-bench. The release broadens competition with Google, Amazon and Datadog models and offers open weights under a modified MIT license, while some rival comparison figures are not independently matched on vendor pages.