The Pulse
Mistral Releases Timestral Forecasting Models as Small as 5M Parameters
Mistral released Timestral, a family of four time-series forecasting models ranging from 5 million to 209 million parameters. The company says the models can make probabilistic forecasts on CPU and rank first on three public benchmarks amon
AI.info Team · Written with AI ·
Mistral says its smallest Timestral model can forecast one day ahead from three weeks of 15-minute data in 10 milliseconds on an eight-thread server CPU. Its largest model takes 140 milliseconds on the same task, according to the company. The figures accompany the release of Timestral, a family of time-series forecasting models announced on October 15, 2026.
The four models have 5 million, 28 million, 123 million and 209 million parameters. Mistral is pitching them for forecasts built from operational data such as equipment sensors, energy use, supply-chain activity and financial flows—work that can involve many related time series and uncertainty about future values.
Four model sizes, one forecasting family
Timestral produces probabilistic, multivariate forecasts: it predicts related series together and returns a range of possible outcomes rather than only one value. It accepts past-only inputs as well as known-future information, such as production schedules or planned shutdowns. The model supports contexts of up to 32,768 time steps and outputs 21 quantiles for each forecast step.
That combination is aimed at problems where a forecast depends on several signals at once. A plant, for example, might use vibration, temperature and pressure readings together; a scheduled production change could also be supplied as information known in advance. Mistral says Timestral can run without a GPU, and that the smaller models trade some accuracy for speed.
Mistral’s benchmark claim comes with a caveat
Mistral says the 209-million-parameter model ranks first among pretrained, zero-shot models on GIFT-Eval, TIME and fev-bench. The company writes, “Our largest model, at just 209M parameters, ranks first on all three of the field's main public benchmarks.” It says those results use one checkpoint without ensembling; its announcement also says results have been submitted to the GIFT-Eval and TIME leaderboards.
The comparisons are the company’s reported evaluation, not an independent assessment in the announcement. Mistral says its 28M model beats Chronos-2, which it describes as about four times larger, on GIFT-Eval and TIME; the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. A technical report with full results is still to come, leaving external readers to examine the detailed methods and comparisons when it appears.
Weights and code carry different licenses
Mistral has published a GitHub repository for running Timestral on users’ own data and links to model weights. The repository code is under Apache 2.0, while the weights use a Modified MIT license, according to the project materials. Mistral also says organisations can adapt the model through its Forge training pipeline, including fine-tuning on proprietary data and deploying the pipeline on their own hardware, including air-gapped environments.
The release puts a compact, general-purpose forecasting model alongside a practical route for organizations to test it on local data. Its usefulness will depend on whether its benchmark performance holds across real deployments, and how much tuning different industries need. For now, Mistral’s published CPU example is specific: a one-day forecast from three weeks of quarter-hour readings, completed in 10 milliseconds by the 5M model.