Gryphe/Codex-24B-Small-3.2

Gryphe/Codex-24B-Small-3.2: A Research-Oriented Roleplay Model

Codex-24B-Small-3.2 is a text-only model developed by Gryphe, built upon the Mistral-Small-3.2-24B-Instruct-2506 architecture. This model represents a research-oriented experiment focused on inducing synthetic diversity in roleplay and storytelling, specifically designed to overcome typical AI response patterns like repetitive phrasing.

Key Capabilities & Features

• Diverse Storytelling: Engineered to embrace a full spectrum of human storytelling, from wholesome to dark, without judgment, aiming for highly varied narrative outputs.
• Repetition Reduction: Leverages the base model's repetition reduction capabilities and Gryphe's systematic pattern diversity research to ensure varied sentence structures and response beginnings.
• Curated Training Data: Fine-tuned on approximately 39 million tokens, combining:
  • GPT 4.1 Instruct core for clean instruction following.
  • DeepSeek V3/R1 roleplay data.
  • Curated "best of" Pantheon interactions.
  • Diverse text adventure compilations.
• Structural Variance: Datasets were specifically validated for structural variance, featuring diverse sentence patterns and 10-40 turn conversations to break common AI response patterns.
• Flexible Narrative Styles: About half of the roleplay data uses Markdown asterisk format, while the majority of other data is in narrative (book-style) present tense, second-person perspective.

Recommended Use Cases

This model is particularly well-suited for:

• Creative Roleplay: Generating highly diverse and non-repetitive responses for interactive roleplaying scenarios.
• Text Adventures: Powering dynamic and varied text-based adventure games.
• Experimental Narrative Generation: Researching and developing AI systems that produce more human-like and less predictable narrative content.

Users are encouraged to provide feedback to the developer on model performance.