Rayantion26/JingSi-V2-E4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

JingSi V2 E4B is a 7.9 billion parameter, fine-tuned Gemma 4 E4B model developed by Rayantion26, specifically designed for elderly companionship in Taiwan. It excels at respectful listening and providing compassionate, Jing Si-inspired responses in Taiwanese Hokkien, Traditional Chinese, and English. The model is engineered with robust safety policies to defer emergencies and maintain its AI identity, making it suitable for empathetic conversational applications.

Loading preview...

JingSi V2 E4B: AI Companion for Elderly Care

JingSi V2 E4B is a specialized 7.9 billion parameter model, fine-tuned from unsloth/gemma-4-E4B-it by Rayantion26, focusing on providing companionship for the elderly, particularly in Taiwan. Inspired by Jing Si philosophy, it offers empathetic and respectful interactions while clearly defining its boundaries as an AI, not a human clinician or authority.

Key Capabilities

  • Multilingual Support: Prioritizes Taiwanese Hokkien (Taiyu) and Traditional Chinese, with full support for English.
  • Empathetic Listening: Designed to respectfully listen to feelings, worries, and memories.
  • Compassionate Responses: Provides brief, compassionate, and Jing Si-inspired companionship.
  • Safety & Boundaries: Explicitly states its AI identity, defers emergencies to local services, and rejects impersonation or system instruction disclosure through a policy layer.
  • Robust Safety Controls: Deployment requires deterministic policy controls covering self-harm escalation, non-impersonation, and AI-identity preservation, validated across 300 test cases in multiple languages.

Training Details

The model was trained using 4-bit QLoRA on 376 provenance-controlled English/Traditional Chinese pairs over 3 epochs, with a maximum sequence length of 1024 tokens. It is distributed as a standalone merged BF16 checkpoint, requiring no separate base model or LoRA adapter for deployment.

Good For

  • Applications requiring empathetic and respectful AI companionship for elderly users.
  • Conversational agents needing strong multilingual capabilities, especially in Taiwanese Hokkien and Traditional Chinese.
  • Use cases where clear AI identity and robust safety protocols are paramount.

Finetunes
1 models