Release notes
In 3.10.1, a turn could stop halfway with "Model output hit its output-token limit" as soon as a screenshot was in the chat. This release gives replies their full length back.
Fixed
- Screenshots no longer cut replies short. Empryo counted every byte of a picture as text, so one screenshot made the chat look full and each reply was held to about a thousand tokens. The model's thinking used that up, and the turn ended without an answer, even after it tried to continue. A picture now counts for the couple of thousand tokens it really costs.