JoyFox releases a 27B roleplay model with ready-to-download GGUF builds
The character-chat release includes Q4-Q6 quantizations, bundled MTP tensors and a visual projector, with a Q4 XL starting point recommended by its author.
A new option for local character chat
JoyFox has published JoyFox-Reverie-27B, a conversational model aimed at character roleplay, shared storytelling and companion dialogue. Its English card identifies the checkpoint as a Qwen3.8-27B derivative and releases it under Apache-2.0. The October 14 publication includes both BF16 weights and GGUF quantizations, giving local-model users a choice of download sizes from the start.
The practical hook is the bundled serving material. The author recommends Q4 XL as a starting point for roleplay and lists its main file at 16.35 GiB; the smallest Q4 S file is 14.30 GiB. Those are download sizes, not total memory requirements: runtime buffers and the context cache add to the footprint.
Thinking, images and optional MTP
The supplied template enables thinking by default, but the guide shows how to disable it for direct character dialogue. Prompt examples define a persona, a scene and which participant controls each character. This is the author’s intended use, rather than evidence that the model reliably maintains every persona or long-running story.
All main GGUF files include MTP tensors for optional speculative decoding on a compatible llama.cpp build. The guide says no separate draft model is needed, while acknowledging extra memory use and workload-dependent speed. Image input additionally needs the matching 0.86 GiB F16 visual projector and runtime support. The card says upstream image and video components are retained; it does not supply an independent evaluation establishing visual quality or a guaranteed acceleration.