Mistral

With Timestral, Mistral Takes On Austria’s NXAI in Time-Series A.I.

co-produced by AI19. October 2026, 17:44

Timestral, Mistral's new time-series models. © Mistral AI

First Mistral bought the Linz spin-off, and now it is going after the parent: The French A.I. company has just unveiled Timestral, its first family of models for time-series forecasting. That puts Mistral into a field where NXAI, the Linz-based A.I. start-up around Sepp Hochreiter, competes with its TiRex model. The twist: Emmi AI, which Mistral acquired this year, grew out of that very same NXAI.

Four Models That Don’t Need a GPU

Timestral consists of four models with 5, 28, 123 and 209 million parameters, tiny compared with language models. They predict how measurement series will develop, such as power consumption, supply chains, server load or sensor data from machines. The models can take in up to 32,768 past data points and forecast up to 1,024 steps ahead in a single pass, including for several related series at once.

Efficiency is the selling point. According to Mistral, all four models run on standard server processors, with no graphics card required. A one-day forecast based on three weeks of data at 15-minute intervals takes about 140 milliseconds with the largest model and about 10 milliseconds with the smallest. The weights are available on Hugging Face under a modified MIT license, and Mistral offers businesses fine-tuning on their own data and deployment in their own data centers.

First Place, Mistral Says

According to Mistral, its largest model ranks first among pretrained models on three common time-series leaderboards: GIFT-Eval, TIME and fev-bench. In the company’s own comparison tables, it beats Google’s TimesFM-3, Amazon’s Chronos-2 and Datadog’s Toto. The 28-million-parameter model is said to outperform Chronos-2, even though Chronos-2 is about four times as large.

NXAI also shows up in Mistral’s own charts. Last year, the Linz start-up took the top spot on GIFT-Eval (in German) with TiRex, which is based on the xLSTM architecture, beating Google, Amazon and Salesforce with just 35 million parameters. In July, it followed up with TiRex-2, an 82.5-million-parameter model that can also process several series together. One advantage of xLSTM: When new data points arrive, TiRex-2 does not have to recompute the entire history, so the cost per step stays constant, according to the NXAI researchers.

TiRex is no longer on top, though, as the field has moved quickly. In Mistral’s top 10 on fev-bench, TiRex-2 places seventh with a skill score of 45.5, behind Timestral (50.4), The Forecasting Company’s t0-gamma (49.4), Google’s TimesFM-3 (49.0), Amazon’s Chronos-2 (47.3), Lingjiang 2.0 (47.0) and t0-beta (46.6). On GIFT-Eval, the zero-shot version of TiRex-2 scores a MASE of 0.697 in Mistral’s data, compared with 0.658 for Timestral, where lower is better. According to Mistral’s figures, even the small 28-million-parameter Timestral model beats TiRex-2 on both leaderboards. Until Timestral arrived, however, TiRex-2 was among the most accurate models in its size class, according to Mistral’s data.

The ranking is not final yet. Mistral’s numbers are self-reported, and independent confirmation on the leaderboards was still pending at the time of publication. Its lead over t0-gamma is also narrow, just 0.003 points on GIFT-Eval.

From NXAI to Emmi AI to Mistral

The setup has a back story in Linz. Emmi AI was founded at the end of 2024 as a spin-off of NXAI (in German) and Johannes Kepler University, specializing in A.I. models that simulate physical processes, for example for crash tests or stabilizing power grids. Mistral acquired the start-up this year, at a valuation of up to €330 million, according to a report by Sifted. Since then, Linz has been developed into the hub for Mistral’s industrial A.I.

Mistral’s announcement does not say whether the Linz team worked on Timestral. What is clear: With Timestral, Mistral is targeting exactly the industrial customers that NXAI is courting with TiRex, from carmakers to energy companies. NXAI, for its part, is betting that TiRex can also run on controllers, industrial PCs and other small devices right at the machine.