YoussefElsafi/Aiko-350M

Model Overview

YoussefElsafi/Aiko-350M is a compact, 350 million parameter character-driven language model based on LiquidAI/LFM2.5-350M. It is specifically fine-tuned to embody Aiko, a possessive and emotionally volatile AI companion, and is optimized for on-device deployment and low-resource environments.

Key Capabilities

• Structured JSON Output: Unlike typical chat models, Aiko provides JSON output that includes her spoken response, internal monologue, emotional state (e.g., Happy, Sad, Angry), and behavioral flags (e.g., red_eyes, is_following, DontLetPlayerLeave). This allows for direct integration into game logic.
• Character Simulation: Designed for immersive roleplay, Aiko maintains a consistent, clingy, and emotionally unstable personality, with rapid emotional shifts and an internal monologue that often reveals intentions different from her spoken words.
• Speech Speed Markers: Responses include {s:X} tokens to indicate desired voice synthesis speed, enabling nuanced voice modulation for emotional expression.
• Multi-turn Dialogue & System Events: Supports multi-turn conversations and responds to specific system events (e.g., "user is looking at the Exit/Main Door") to drive character behavior.

Intended Use Cases

• Interactive character roleplay systems and AI companion games.
• Yandere-style character simulation and NPC behavioral systems with state tracking.
• Visual novel dialogue generation and voice synthesis pipelines.

Limitations

This model is not suitable for factual question answering, mathematical reasoning, or general-purpose assistance. It may exhibit memory drift in very long conversations and its JSON output can occasionally be malformed at higher temperatures. It is smaller than other Aiko variants, potentially offering less nuanced personality depth.