Timestral
A family of time-series foundation models released by Mistral, based on a zero-shot forecasting paradigm. It supports CPU inference, can output multi-quantile probabilistic forecasts, and accepts covariate inputs.
Timeline (last 90 days)
Mistral launched Timestral, its first time-series foundation model family, with four sizes (5M, 28M, 123M, and 209M), running purely on CPU, with per-forecast latency of 10–140ms and support for a 32768-step context.
Mistral says the 209M version of Timestral ranks first among zero-shot pretrained models on GIFT-Eval/TIME/fev-bench.
The largest Timestral version (209M parameters) ranks first on three major public benchmarks.
Timestral does not require a GPU for inference.
The Timestral weights are released under a modified MIT license, with the restriction that a company is not authorized to use them if its global monthly revenue last month exceeded $20 million.
The Timestral inference code is under the Apache 2.0 license, while the weights use a modified MIT license.
Mistral released Timestral, a family of four time series foundation models (5M, 28M, 123M, 209M) sharing the same architecture and training recipe.
The Timestral documentation lists Timestral 1.0 as a public preview and does not state API pricing.
Installing Timestral requires installing from GitHub, requires Python 3.10+ and torch 2.4+, and the package is not on PyPI.
Mistral positions Timestral for local, VPC, and air-gapped deployments.
3 more timeline events
All Facts (13)
Mistral says the 209M version of Timestral ranks first among zero-shot pretrained models on GIFT-Eval/TIME/fev-bench.
50%UnverifiedThe Timestral models are zero-shot, so users can make predictions without training on their own data.
50%UnverifiedThe largest Timestral version (209M parameters) ranks first on three major public benchmarks.
50%UnverifiedTimestral does not require a GPU for inference.
50%UnverifiedThe Timestral weights are released under a modified MIT license, with the restriction that a company is not authorized to use them if its global monthly revenue last month exceeded $20 million.
50%UnverifiedThe Timestral inference code is under the Apache 2.0 license, while the weights use a modified MIT license.
50%UnverifiedThe Timestral documentation lists Timestral 1.0 as a public preview and does not state API pricing.
50%UnverifiedInstalling Timestral requires installing from GitHub, requires Python 3.10+ and torch 2.4+, and the package is not on PyPI.
50%UnverifiedMistral positions Timestral for local, VPC, and air-gapped deployments.
50%UnverifiedTimestral supports inputting multiple series and covariates (known in the past or known in the future) at the same time, with missing values passed in as NaN.
50%UnverifiedMistral says Timestral's results come from a single checkpoint, with no ensemble, and have been submitted to the GIFT-Eval and TIME leaderboards, and that most of the training data is synthetic.
50%UnverifiedMistral launched Timestral, its first time-series foundation model family, with four sizes (5M, 28M, 123M, and 209M), running purely on CPU, with per-forecast latency of 10–140ms and support for a 32768-step context.
50%UnverifiedMistral released Timestral, a family of four time series foundation models (5M, 28M, 123M, 209M) sharing the same architecture and training recipe.
50%