Mistral has introduced Timestral, its first foundation model series built for time-series tasks. The lineup includes four models ranging from 5 million to 209 million parameters. According to the announcement, the largest 209M-parameter model ranked first on all three public benchmarks cited by the company: GIFT-Eval, TIME, and fev-bench. Mistral said the model outperformed existing alternatives with larger parameter counts.
Timestral supports probabilistic multivariate forecasting and works with both past and future covariates. The company also said the models can run efficiently on CPUs and support zero-shot forecasting. On the technical side, Timestral combines synthetic pretraining, covariate-aware attention, and long-context learning. Mistral is also offering domain-adaptation fine-tuning through Mistral Forge.
The company disclosed CPU inference figures for two model sizes on a standard 8-thread server CPU. The 5M model takes 10 milliseconds to forecast the next day, while the 209M model takes 140 milliseconds.
Mistral has released Timestral, its first foundation model series for time-series forecasting. The series includes four models ranging from 5 million to 209 million parameters.
The largest version, with 209 million parameters, ranked first across three public benchmarks: GIFT-Eval, TIME, and fev-bench. Mistral said it outperformed existing models with larger parameter counts.
Forecasting features and deployment
Timestral supports probabilistic multivariate forecasting and is compatible with both past and future covariates. The company said the models can run efficiently on CPUs and support zero-shot forecasting.
Training design and fine-tuning
The model combines synthetic pretraining, covariate-aware attention, and long-context learning. Mistral is also offering domain-adaptation fine-tuning through Mistral Forge.
CPU inference speed
On a standard 8-thread server CPU, the 5M model takes 10 milliseconds to forecast one day ahead, while the 209M model takes 140 milliseconds.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
