Mistral AI released Timestral on October 15, 2026, its first family of time-series foundation models. The lineup ships in four sizes: 5M, 28M, 123M and 209M parameters. The company says the largest checkpoint ranks first on the field's three main public benchmarks, GIFT-Eval, TIME and fev-bench, while using only about two-thirds the parameters of the previous state of the art it cites.
Timestral targets probabilistic multivariate forecasts and can condition on past covariates or on known future ones such as calendars and production schedules. Mistral places each size on the accuracy-versus-size frontier and says no public peer at that footprint beats it. Leaderboard entries were filed for GIFT-Eval and TIME.
Training blends synthetic pre-training with attention that can read covariates and with long context windows. Mistral says CPU inference stays efficient and zero-shot transfer helps when a new domain has little labelled history. Domain adaptation is offered through Mistral Forge fine-tuning rather than forcing every team to train a forecaster from scratch.
Enterprise motivation is practical. Sensor telemetry, observability metrics, energy loads, supply-chain flows and financial volumes are all series that drive planning decisions. Classical tools such as ARIMA still fit a single series well, and specialised deep models such as PatchTST learn across related series, but they usually do not transfer zero-shot to a new domain the way a pre-trained foundation model aims to.
On GIFT-Eval and TIME, Mistral reports a 28M checkpoint topping Chronos-2 at roughly one-quarter the size, while the 209M version surpasses TimesFM-3 and Falcon-2.0 despite a smaller parameter count. Those comparisons are vendor-reported leaderboard claims from a single checkpoint without ensembling; this package treats them as attributed results, not independently re-run scores.
Availability details in the launch post emphasise CPU-friendly inference and a shared architecture across the four sizes. Smaller models trade a bit of accuracy for speed and cost; larger ones push forecast quality. Mistral presents the family as a way for industrial and financial teams to start from a general prior instead of fitting a bespoke network for every plant or product line.
Mistral also stresses that every size shares the same architecture and training recipe, so teams can move up or down the family without redesigning prompts or fine-tuning pipelines. That matters for factories and finance desks that want a small CPU model on the edge of a plant and a larger sibling in the central planning stack.
Open questions after the announcement include exact download and API packaging, licensing terms beyond the blog summary, and how quickly third-party leaderboard mirrors will confirm the submitted rankings. Even with those caveats, a first-party time-series foundation family from a major European model lab is a concrete product event for builders who live in telemetry and planning data rather than chat prompts.