Changelog

Chat Improvements

• Added syntax highlighting to the input bar for better visual clarity
• Implemented autocomplete for {{char}} and {{user}} variables
  • Triggers automatically when typing { character
• Added plus button on chatbar to quickly add {{user}}, {{char}}, and ooc:
• Applied OOC color syntax highlighting to {{char}} and {{user}} variables for better readability
• Added "Regenerate with Instruction" feature for more control over responses
• You can now set a custom chat title for better organization

Character Discovery

• Smarter default sorting based on user status:
  • Anonymous and new users (accounts < 72 hours): Default to "Top"
  • Regular users: Default to "Hot" for more relevant content
• "New" tab improvements: Now sorted by publication time instead of creation time
• Imported bots deprioritized in favor of manually created bots for better quality

UI Polish

• Refined minor UI details

Bug Fixes

• Fixed keyboard overlap issues with bottom drawer memory/custom instruction sections on mobile
• Fixed featured characters not showing tags / blacklist tags not working on featured characters

Fantasia 2 Released

• Completely new model - Order of magnitude larger than Fantasia 1, Stheno, or Magmell
• Significantly improved intelligence

Intelligence Improvements

• Better context understanding - Characters react realistically and notice subtle details
• Does not miss obvious cues or misinterpreting situations
• Characters understand implications and respond appropriately
• More coherent handling of complex scenarios

Writing Quality

• Intentional prose - Descriptions serve the scene, not just filler
• Eliminated unnecessary flowery language that added nothing
• Less purple prose for its own sake

OOC (Out of Character) Instructions

• Simple format - Just use ooc: <instruction> (easiest to type)
• Also supports OOC:, (OOC:), and [OOC:] formats
• Better instruction following than older models
• Examples of what you can do:
  • Guide romantic scenes: ooc: make them more receptive
  • Character development: ooc: have them drop their guard
  • Time control: ooc: time skip one year
  • World changes: ooc: transform into cyberpunk dystopia
  • Power fantasy: ooc: everyone except us freezes in time

Generation Improvements

• Better regenerations - More diverse responses instead of slight rewording
• Story continuation - Continue button actually progresses the story
• Works well with OOC guidance to avoid loops
• No more "why aren't you saying anything?" spirals

NSFW Enhancements

• More accurate anatomical descriptions
• Better understanding of emotional context

Reduced Issues

• Minimal impersonation - Only 1 occurrence in 400+ test messages
• Much less confusion about who's speaking or acting

Known Issues

• Some phrase repetition common to AI models ("breath hitching", "barely above a whisper")
• Occasional context confusion in very long conversations
• Writing quality can degrade with minimal effort responses without OOC guidance

Model Updates

• The exp01 model has been renamed to Fantasia and is now the primary model
• Fantasia is no longer experimental - it's now the main model
• Stheno and Magmell models still available for those who prefer them
• Significant optimizations to Fantasia's efficiency through extensive team effort
• More efficient inference engine enables major memory improvements

Memory Improvements

• Context length doubled from 6k to 12k tokens for Fantasia (holds ~9,000 words of conversation history)
• Character token limit increased from 2k to 3k tokens (thanks @Dante for the suggestion!)
• Future goal to increase character token limits even further
• Long term memory improvements will continue beyond context length

Response Improvements

• Fixed over-reliance on character description and first message (less repetition)
• More responsive to current conversation context, not latching onto greeting messages
• Continue now generates more diverse responses (interprets as persona staying silent)
• Reduced impersonation occurrences significantly

Fixes

• Fixed italics appearing inside quotes
• Resolved blockquote mode showing two blockquotes on the same line
• Character image uploads no longer appear blurry
• Various other minor fixes