📈Mistral Unveils Timestral: 209M Param Time-Series Models
Timestral's 209M model beats bigger rivals on accuracy
TL;DR
Mistral's Timestral models, ranging from 5M to 209M parameters, deliver state-of-the-art forecasting accuracy. The largest model outperforms rivals like Chronos-2 and Falcon-2.0 despite having fewer parameters.
Mistral has introduced Timestral, a family of time-series foundation models with sizes from 5M to 209M parameters. The largest model, at 209M parameters, outperforms larger rivals like Chronos-2 and Falcon-2.0 on accuracy benchmarks. Timestral's smaller models offer faster inference and lower costs, making them ideal for real-time applications. The models are trained on diverse data and can forecast one day ahead from three weeks of 15-minute history on a standard server CPU.
Key Points
Timestral's largest model has 209M parameters, outperforming rivals like Chronos-2 and Falcon-2.0.
The models range from 5M to 209M parameters, optimized for every specification.
Timestral can forecast one day ahead from three weeks of 15-minute history on a standard server CPU.
The 28M model outperforms Chronos-2, about four times larger, on GIFT-Eval and TIME leaderboards.
Timestral uses decoupled horizon forecasting, covariate-aware attention, and synthetic-first data for training.
Why It Matters
If you're building a forecasting system for industrial data, Timestral's multivariate forecasting and probabilistic capabilities can improve forecast quality. The 209M model outperforms larger rivals, while the 28M model offers better performance than Chronos-2 despite being smaller. Timestral's ability to run on CPUs and handle zero-shot forecasting makes it a strong choice for diverse applications.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,566 builders reading daily.