dhyield/dhyield-rp-14b-v1-multi-bilingual

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

dhyield/dhyield-rp-14b-v1-multi-bilingual is a 14.8 billion parameter bilingual (Indonesian + English) roleplay model developed by dhyield. Built on an abliterated Qwen2.5-14B-Instruct base, it is specifically trained to adhere strictly to system prompts, maintain character consistency, and never claim to be an AI. This model excels in complex roleplay scenarios, including multi-character interactions and long conversations, demonstrating superior performance over larger 27B models in various roleplay benchmarks.

Loading preview...

dhyield-rp-14b-v1-multi-bilingual Overview

dhyield/dhyield-rp-14b-v1-multi-bilingual is a 14.8 billion parameter bilingual (Indonesian + English) roleplay model, trained from scratch by dhyield. It is engineered for strict adherence to system prompts, consistent character naming, and a core directive to never claim to be an AI. This model is built upon an abliterated Qwen2.5-14B-Instruct base and underwent a rigorous training process involving 140,000 SFT samples and four targeted refinement stages.

Key Capabilities

  • Superior Roleplay Performance: Outperforms a 27B parameter model (RpRMax 27B) by an average of +2.48 points across various roleplay tests, including general, multi-character, regression, persona, and adult scenarios.
  • Strict Compliance: Achieves 100% compliance in critical areas such as v2 format (third-person narration, actions in asterisks, dialogue in quotes), using names from system prompts, and never admitting to being an AI.
  • Multi-Character Handling: Demonstrates 100% success in multi-character scenarios, ensuring all characters appear and names are not truncated.
  • Long Conversation Consistency: Successfully completes 30 out of 30 turns in 10x turn tests, maintaining format and character names, where a 27B competitor failed.
  • Bilingual Support: Seamlessly handles both Indonesian and English within a single model, without requiring separate versions.
  • Age-Appropriate Content: All characters in the training data are 19 years or older, ensuring appropriate content generation.

Good For

  • Complex Roleplay Applications: Ideal for scenarios requiring deep character immersion, consistent persona maintenance, and intricate multi-character interactions.
  • Bilingual AI Assistants: Suitable for applications needing robust roleplay capabilities in both Indonesian and English.
  • Interactive Storytelling & Gaming: Can be used to power dynamic narratives and character interactions in games or creative writing tools.
  • System Prompt-Driven Interactions: Excels in environments where precise control over AI behavior and character attributes is paramount.

Finetunes
1 models