You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This version of Icegrams is trained on roughly 10x the token count, around 1B tokens spanning texts from 1980 to mid-year 2026, with a gentle recency bias. Miðeind's neural corrector, Málfríður, has been applied to every sentence to clean up noise in the input text.
The resulting model's size is 213.2MB, compared with the old model's 43MB.