Model Selection

• Rochat-Character: A hybrid model combining modeling and engineering techniques. To ensure the basic experience of Rochat Character (including proficiency in role-playing and storytelling, having sufficiently long memory at the model level, good reasoning ability, and a wide knowledge base), we employ some engineering techniques. In the Rochat-Character option, we integrate the GPT-3.5 Fine Tuning model, Claude 3 model, and our own 13B model, which is fine-tuned based on open-source models and role-playing scenarios for character output reasoning. This approach ensures a decent experience in every aspect and avoids restrictions imposed by major companies, enabling the generation of NSFW content. This technical solution is still evolving, and we are currently aware of output instability.

• Rochat-RP: This model is specifically designed for original characters, stories, and NSFW content support. It is a 13B model to which we have added some Lora components. It is more proficient in English content.

• GPT-4-Turbo/GPT-4o: The most powerful model released by OpenAI, offering faster output speeds and better suitability for role-playing scenarios and real-time reasoning.

• Claude 3/3.5: The most powerful model released by Anthropic, improving character memory experience through enhanced support for long texts. It aims to make the user experience more enjoyable and optimizes for various cognitive tasks such as Q&A, text generation, conversation, etc.

• Rochat-Llama 3-70b: As one of the best role-playing models currently on the market, Rochat-LLama 3-70B performed excellently in our tests, ranking among the top mainstream models, and surpassing some performance indicators of GPT-4o and the latest Claude-3.5 Sonnet models. It supports a context length of 128k, enabling better handling of ultra-long conversations. With 70B parameters, it enhances model consistency and improves response diversity while also being very proficient at handling NSFW content.

• Rochat-Llama 3-70b-chat: Rochat-LLama 3-70B-Chat is a model based on and extended from Rochat-LLama 3-70B, featuring ultra short responses and wild, off-the-wall writing. Simple with no gimmicks, requires that {char} is a character.

• Gemini 1.5:

  • Long Memory: Outstanding performance in long-context scenarios.

  • Rich Knowledge Base: The model itself has a sufficiently rich knowledge base, providing an excellent interactive experience with relatively few prompts for most types of characters.

  • Exceptional Multilingual Ability: Gemini excels not only in English but also shows outstanding support for other languages such as Japanese, Chinese, Spanish, French, German, and Arabic.