Open sourceTimestral: Mistral AI’s first time-series forecasting models
Mistral AI has released Timestral, four time-series forecasting models from 5M to 209M parameters that the company says run without a GPU. The 209M weights’ license does not apply to companies with more than $20 million in monthly revenue.
Source of truthmistral.ai/news/timestralOn October 15, 2026, Mistral AI released Timestral, its first family of time-series forecasting models, in four sizes from 5M to 209M parameters. According to Mistral, it needs no GPU: on a server CPU with 8 threads, a one-day forecast from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M. Mistral also says the 209M ranks first on GIFT-Eval, TIME and fev-bench; a technical report will follow. Fine-tuning runs through Mistral Forge. The 209M weights’ “Modified MIT” license does not apply to companies with more than $20 million in revenue the previous month.
Other sources: github.com/mistralai/timestralhuggingface.co/…/LICENSE
Related stories
Granite Speech 5.0: IBM releases Apache 2.0 speech models
IBM released granite-speech-5.0-2b and revised two more speech models, all under the Apache 2.0 license. Greek is not listed for the 2B and 470M models.
Beam: Reflection AI’s open 501B model under Apache 2.0
Reflection AI has published the weights of Beam, a 501B-parameter model, on Hugging Face under Apache 2.0, which also allows commercial use.