Mistral Launches Timestral Time-Series Model Family, Taking Aim at TimesFM-3

Mistral AI released Timestral on October 19, its first family of time-series foundation models. The Hugging Face repository for the flagship Timestral-209M is dated October 15, suggesting the weights went live a few days before the formal announcement.

Timestral is not a chat or coding model. It takes an ordered sequence of numbers — daily retail sales, hourly power demand, a stock's trading volume — and forecasts what comes next. Mistral says it supports zero-shot forecasting, probabilistic outputs, multivariate prediction and both past and future covariates, meaning it can handle an unseen dataset without retraining, express uncertainty and account for external drivers such as holidays or weather.

The family ships in four sizes: 5 million, 28 million, 123 million and 209 million parameters. That is more than two orders of magnitude smaller than the hundreds-of-billions-scale language models that dominate 2026 releases. Mistral's argument is that time-series data carries far less entropy than natural language, so specialised small models do not need the same parameter budget.

Benchmarks are the centrepiece of the launch. Mistral reports that Timestral-209M scored a MASE of 0.658 and a CRPS of 0.446 on GIFT-Eval, 0.626 and 0.522 on TIME, and a SQL skill score of 50.39 on fev-bench. Google's TimesFM-3 is listed at 0.667, 0.456, 0.640, 0.536 and 48.99 respectively. Mistral says the 209M model leads all three benchmarks using a single checkpoint, without ensembling.

The crown is contested. Google's own TimesFM-3 announcement claims the top spot across GIFT-Eval, fev-bench and TIME, covering both point and probabilistic metrics. Two vendors, overlapping benchmark suites and opposite conclusions: until independent researchers reproduce the comparison under matched conditions, both sets of numbers remain vendor-reported.

Mistral also claims that its 28-million-parameter variant beats the competing Chronos-2, a model roughly four times larger, on GIFT-Eval and TIME. If that holds, it may matter more than the flagship win, because it suggests the efficiency gain is not a one-off property of the largest checkpoint.

Commercially, forecasting is one of the most directly monetisable AI applications, with demand across finance, retail, logistics and industrial operations. Small models also change deployment economics: third-party write-ups describe Timestral-209M as CPU-capable, though Mistral has not confirmed latency or memory requirements. For companies running thousands of forecasts a day, cost per prediction may matter more than a few thousandths of a point on a leaderboard.

Pricing, a hosted API, detailed licensing and self-hosting requirements remain unconfirmed. How Mistral converts open weights into enterprise revenue is therefore still an open question. The next signals to watch are independent reproductions, Google's response and how Timestral performs on real business data.