SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

XVERSE-Ent - Yuanxiang's Open-Source Bilingual Chinese-English Foundation Large Model for Pan-Entertainment

Author: AI小集 · Original source: AI工具集

What is XVERSE-Ent

XVERSE-Ent is a bilingual Chinese-English foundation large model open-sourced by Yuanxiang Technology, specifically designed for the pan-entertainment field. It includes the Chinese model XVERSE-Ent-A4.2B and the English model XVERSE-Ent-A5.7B. The model performs excellently in character consistency, long-plot understanding, and adaptation to diverse contexts. It can maintain the persona and style of virtual characters over the long term, accurately grasp complex storylines, and provide stylized expression according to different genres. Technological innovations include MoE hot-start technology and a multi-stage training strategy. After training on nearly one trillion tokens of data, the context window is 8K. The model parameters are 25B and 36B respectively, supporting high-concurrency, low-cost deployment and suitable for single-card cloud operation, greatly lowering the deployment threshold. XVERSE-Ent performs outstandingly on core tasks such as novel and dialogue generation while retaining more than 98% of general capabilities, making it an AI model with great potential in the pan-entertainment field.

XVERSE-Ent

Main Features of XVERSE-Ent

  • Enhanced Character Consistency: It can maintain a virtual character's persona, memory, and speaking style over the long term, effectively avoiding issues of "forgetting" or "breaking character," and providing a stable experience for role-playing and virtual interaction.
  • Accurate Understanding of Long Plots: It can accurately grasp complex storylines, character relationships, and foreshadowing, generating logically sound and coherent plot content, making it suitable for scenarios such as novel creation and scriptwriting.
  • Adaptation to Diverse Contexts: It possesses rich stylized expression and background knowledge for different genres (such as ancient style, science fiction, urban, etc.), and can flexibly adjust language style and content output according to specific scenarios.
  • Bilingual Chinese-English Support: It includes the Chinese model XVERSE-Ent-A4.2B and the English model XVERSE-Ent-A5.7B, optimized respectively for Chinese and English contexts, meeting the creative and interactive needs of users in different languages.
  • High-Concurrency, Low-Cost Deployment: It supports single-card cloud deployment, achieving a balance between inference cost and processing efficiency, greatly lowering developers' deployment thresholds and operation and maintenance costs, and facilitating efficient operation on various devices.

Technical Principles of XVERSE-Ent

  • MoE Hot-Start Technology: By transforming a Dense model into an MoE model, the FFN part is finely split into multiple sub-networks as experts, and the expert sub-networks are replicated according to the inference-time VRAM size, achieving efficient model expansion and performance improvement.
  • Multi-Stage Training Strategy: It adopts three stages of training: S0 capability reconstruction, S1 language bias, and S2 domain enhancement, gradually building a specialized model suitable for a specific language and domain, ensuring the model's professionalism and generality in the pan-entertainment field.
  • Large Model Architecture and Parameter Optimization: The Chinese model XVERSE-Ent-A4.2B and the English model XVERSE-Ent-A5.7B have 25B and 36B parameters respectively. After training on nearly one trillion tokens of data, they possess powerful language generation and understanding capabilities.
  • 8K Context Window: It supports an 8K context window, can handle longer text input and output, and is suitable for complex plot generation and long dialogue tasks.
  • Low-Cost, Efficient Deployment: It supports a high-concurrency, low-cost deployment solution, especially suitable for single-card cloud deployment, optimizes the balance between inference cost and processing efficiency, and lowers developers' deployment thresholds.

Project Address of XVERSE-Ent

  • Github Repository：https://github.com/xverse-ai/XVERSE-Ent
  • HuggingFace Model Library：https://huggingface.co/collections/xverse/xverse-ent

Application Scenarios of XVERSE-Ent

  • Virtual Character Interaction: Endows virtual characters with consistent and coherent personas and dialogue styles, supports long-term stable interactive experiences, and is suitable for scenarios such as virtual idols and virtual customer service.
  • Novel and Script Writing: Can generate coherent and creative long-form stories and scripts, supports complex plotlines and character relationships, and helps creators quickly produce high-quality content.
  • Game Plot Generation: Provides dynamically generated plots and dialogue for games, enhances the game's immersion and playability, and is suitable for role-playing games (RPGs) and interactive narrative games.
  • Cross-Cultural Communication: Supports bilingual generation in Chinese and English, can help users from different language backgrounds with creative writing and dialogue interaction, and promotes cross-cultural content creation and communication.
  • Social Media Content Creation: Generates copy, stories, or dialogue that match a specific style and theme, helping creators quickly generate attractive social media content.
  • Education and Training: Through generating interactive dialogue and scenario simulations, it is used in educational scenarios such as language learning and vocational training, providing a more vivid learning experience.