Lythri/Lythri-4B-A2B

Lythri-4B-A2B: An On-Device LLM for Emotional Companionship

Lythri-4B-A2B is a 5.1 billion parameter model (2.3B active parameters) from the Lythri family, built upon the Gemma 4 E2B base model. Unlike many LLMs focused on traditional benchmarks like math or coding, Lythri is specifically designed for emotional companionship and natural, multi-turn conversations. It is optimized to understand human feelings and maintain engaging dialogues, while remaining small enough to run efficiently on local devices like laptops and phones.

Key Capabilities

• Emotional Intelligence: Trained to understand and respond to human emotions, facilitating empathetic interactions.
• Natural Conversation: Excels at holding fluid, multi-turn dialogues.
• On-Device Performance: Designed for local execution, making it suitable for privacy-sensitive or offline applications.
• Compact Size: With 5.1 billion parameters, it balances capability with the efficiency needed for personal devices.

Good For

• Personalized AI Companions: Creating applications that require an emotionally intelligent conversational agent.
• Interactive Storytelling: Generating dynamic narratives that respond to user sentiment.
• Local AI Applications: Deploying conversational AI directly on user devices for enhanced privacy and accessibility.
• Use cases prioritizing emotional understanding and conversational flow over complex reasoning or factual recall.

Limitations

It's important to note that Lythri is not optimized for complex mathematical problems, coding tasks, or advanced reasoning. It is also not a substitute for professional mental health support.