0.5.7
We're happy to announce the release of Soji, a 671B-parameter finetune of DeepSeek V3 0324, at 60K context for Mars subscribers. When you switch to it, you'll be prompted to switch to a popular preset -- please do this for best results, as it has different ideal settings than other models. Non-subscribers will have a few trial messages to try it out, as well.
General notes on settings:
If it's being repetitive, the repetition, presence, and/or frequency penalties are too low.
If it's dropping common words like 'the', 'an', and pronouns, those penalties are too high.
Other updates in this release:
We are live on the iOS App Store again.
"AI Ratings" are being redone by Soji, and the "Underrated" tab will now cycle out cards over time.
In multi-character/user chats, lorebooks for the character/user not being prompted for will no longer be pulled in.
Fixed a bug with anonymized character and user IDs not populating correctly in multi-user/character stages.
You can change the generated captions on gallery images.
Fixed a bug with “Show More” not showing correctly in some browsers.
When switching models, some will now ask if you want to pull in the settings of the most popular preset for that model.
Fixed a bug with templated textgen requests coming from a stage not populating {{char}} correctly.
Fixed a bug with templated textgen requests from stages not allowing min/max tokens below 50.
Fixed failed imports of cards with empty name fields.
Added Claude 4 to the model list.
Fixed an issue with the prompts for some APIs with special characters not building correctly.
Fixed a performance issue with multiplayer.
Made possible/probable tags alias to their base tag, e.g. “possible rape” -> “rape”.
Last updated