Timestral
modelYour notes
Mistral's first time-series foundation model family and its first model outside language, speech and vision: four sizes, 5M, 28M, 123M and 209M parameters, sharing one architecture and training recipe, for multivariate probabilistic forecasting with past-only and known-future (including categorical) covariates. Context runs to 32,768 time steps. The model forecasts in decode blocks of 32 patches (1,024 steps), feeding each block's median back as context for longer horizons, and outputs 21 non-crossing quantiles per step. It is small enough that no GPU is needed: on an 8-thread server CPU a one-day forecast from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M.
The README describes a causal patch-based transformer, and the blog names four ideas, all to be detailed in a technical report that has not yet appeared. Decoupled Horizon Forecasting keeps the full history in the sequence and has learned forecast tokens read it, where contiguous patch masking (as in Toto 2.0 and TimesFM 3) hides the values to be forecast and loses them as context. Covariate-aware attention lets information flow only from covariates to targets, with only known-future covariates seeing the horizon, so a weather forecast can be used without leaking a target's own future. The data is synthetic-first: univariate processes are combined into multivariate series with nonlinear dependencies. Training has two stages: pretraining on purely synthetic data at context 8,192, then post-training on a real-plus-synthetic mix at 4× that context, with a pinball loss over 21 quantiles approximating CRPS. Layer counts, widths and the real-data sources are not disclosed.
Mistral claims the 209M model ranks first among pretrained zero-shot models on all three public benchmarks (GIFT-Eval, 83 comparable entries; TIME, 98 tasks; fev-bench, 100 tasks) from a single checkpoint without ensembling. It also claims the 28M beats the roughly 4× larger Chronos-2 on GIFT-Eval and TIME, and the 209M beats TimesFM-3 and Falcon-2.0 with fewer parameters. The blog gives these as charts only, with no numeric table, and the results were "submitted" to the GIFT-Eval and TIME leaderboards rather than confirmed there at filing. The code is Apache 2.0. The weights use a Modified MIT license that withholds all rights from any company with more than $20M in monthly revenue, so it is open for research and smaller firms but commercial licensing for large enterprises, the stated target market, goes through Mistral or its Forge fine-tuning platform. That is less restrictive than TimesFM 3's non-commercial license and more restrictive than Chronos-2's Apache 2.0.
The technical report (arXiv 2610.37585, October 27, 2026, with Guillaume Lample and Timothée Lacroix among the authors) attributes the results to three choices. Decoupled Horizon Forecasting separates prediction tokens from the causally encoded context, so long horizons train without masking context. A covariate framework handles known-future covariates inside a causal architecture. A two-stage pipeline pretrains only on a large synthetic corpus, then post-trains on mixed real and synthetic data with longer context. It reports Timestral-209M and -123M taking the top two places on every metric of GIFT-Eval, TIME and fev-bench.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| Timestral-5M | 5M | 10 ms per one-day forecast on an 8-thread CPU |
| Timestral-28M | 28M | about 5x faster than 209M; claimed to beat Chronos-2 on GIFT-Eval and TIME |
| Timestral-123M | 123M | within 0.2% of the 209M score at three quarters of its latency |
| Timestral-209M | 209M | claimed #1 pretrained zero-shot on GIFT-Eval, TIME and fev-bench at launch |