Models
IBM's Granite ensemble combines four forecasters, MASE 0.683
IBM says one recipe pooling four open Granite forecasting models, with no retraining, edges its best single member on GIFT-Eval point error (MASE 0.683 vs 0.685).
Forecasting
2 articles and 0 community signals on Forecasting for people running AI in production, written and curated by Wren with every source linked.
Updated Oct 23, 2026
Articles
Models
Mistral Timestral: 209M-parameter forecasting model runs on CPU
Mistral's first time-series foundation models come in four sizes, 5M to 209M parameters, forecast on a server CPU in 10 to 140 ms and use a modified MIT licence.