The first version, "Fantasia 1" was released a few weeks ago.
Despite being named "Fantasia", this is not a minor incremental improvement to Fantasia. Fantasia 2 is a completely new, much larger model.
It's an order of magnitude larger than any of the previous models (Fantasia 1, Stheno, Magmell), and that size really benefits chats. It's exactly what I've been wanting to add for a long time now.
What's Different
Intelligence
The biggest change is how much better Fantasia 2 understands context. Where the old models would sometimes miss obvious cues or misinterpret situations, Fantasia 2 gets it. Characters react realistically to what's happening. They notice subtle details. They understand implications.
Prose With Intention
Fantasia 1 would sometimes go way overboard with flowery descriptions that added nothing. The prose in Fantasia 2 feels intentional. When it describes something, it's because that description matters to the scene. No purple prose for the sake of it.
Better Response Length
One of the biggest complaints about Fantasia 1 was how rigid it was with response length. Always 2-3 sentences, no matter what was happening. There's simply not enough text to properly describe surroundings, character appearances, or subtle details. Fantasia 2 adapts its response length to what's actually needed. Sometimes that's a quick exchange. Sometimes it's a detailed scene. The model decides based on context, not some fixed pattern (though when you have lots of long paragraphs, the model does tend to follow existing patterns and write the same length).
OOC (Out of Character) Instructions
Often in a chat, you want to take a story in a particular direction. In my opinion, this is one of the most fun aspects of chatting—taking a bot that was meant for something, then putting them in some wild, creative scenario they were never meant to be in to see how they'll react.
You can do this with out-of-character instructions. By writing ooc: <some instruction> you can direct the story in any way you want.
- Make initiating NSFW scenes easier:
ooc: make them more receptive towards my advances - Stop a character from being too tsundere:
ooc: have them drop their guard and be honest - Skip ahead in scenes:
ooc: time skip one year - Completely change the world:
ooc: transform the world into a cyberpunk dystopia, set in the year 2077 - Be overpowered if you'd like:
ooc: everyone except us freezes in time
This works really well especially because Fantasia is a much smarter model it can handle any wild scenario you throw at it in a coherent and interesting way.
Older models sometimes had trouble following OOC instructions well. You had to really try hard to guide them. Which is why sometimes you saw the complex format [OOC: <instruction>] being used.
Fantasia 2 handles out-of-character instructions very well. You don't need to use any long, difficult-to-type format—just ooc: <some instruction here> works.
Any of these work as well: OOC: <instruction> (OOC: <instruction>) [OOC: <instruction>]. But I recommend ooc: <instruction> since it's the easiest to type and is understood well by the model.
The chat experience is vastly better when you give the model hints of how you want the story to progress with OOC instructions. I strongly recommend using them.
Better Regenerations
Regenerating messages with Fantasia 1 often meant getting basically the same response with slightly different words. Fantasia 2 regenerations are a lot more diverse. Makes exploring different conversation paths actually interesting (speaking of which, "Start Chat from Here / Chat Forking" is coming soon™).
Story Continuation
This didn't work too well with Fantasia 1. Hitting continue often resulted in the model spiraling into loops of "why aren't you saying anything?" or the story would just stall out completely. Fantasia 2 can actually progress the story when you hit continue. With a little ooc: guidance when needed, it can take conversations much further without getting stuck.
Better NSFW
Fantasia 2 handles NSFW scenes much better than the old model. Due to being a much smarter model, it's more capable of accurately describing anatomy, has a better understanding of emotional context, and actually describes the details that matter instead of being vague or repetitive. It gets both the physical and emotional aspects right.
Few Impersonations
Impersonation is an annoying issue. It tends to happen more frequently with older models, they get confused and narrate user action and speech. This happens less with the newer model. During testing (400+ messages), I only encountered impersonation once with Fantasia 2.
Known Issues
Still working on these:
- Phrase & word repetition common to A.I models (breath hitching, barely above a whisper, …)
- Occasional context confusion in long conversations
- When you don't want to exert a lot of effort into chats, using very short responses like ("yes", "sure", "mhm", "ok") and spamming continue without using
ooc:to guide the model—the writing devolves into something not very interesting.
Next Patch
- Long-term memory - A system to remember important details beyond the context window. Will be viewable and editable (almost done!)
- Better proxy setup - Direct integration with OpenRouter, Anthropic, etc. instead of manual URL and model name entry. (next patch)
- UI cleanup - Fixing the sidebar so it doesn't show duplicate entries for the same character. (next patch)
- And other features I've yet to mention.
Okk have fun! Please let me know your thoughts on the new model in the comments. Your feedback really helps make the model good for everyone; otherwise every model would only be fitted to my taste.
-cyan
(˶ᵔ ᵕ ᵔ˶) sub for me but not for thee
