LongShu · IceBlink

TL;DR: A narrative-aware dialogue generation model specifically built for game NPC interactions, based on GLM-4.5-Air with deep fine-tuning, MLX 4bit quantized, running smoothly on consumer-grade hardware. github：https://github.com/luoyike2003ls/LongShuGameDev

🎮 Model Positioning

General-purpose LLMs often feel "out of character" when generating game NPC dialogue — hollow tone, lack of personality, poor contextual coherence, and unable to support immersive narrative experiences.

LongShu · IceBlink aims to solve this pain point. Built on the Zhipu GLM-4.5-Air foundation, we performed full instruction fine-tuning targeted at game narrative scenarios, creating an AI dialogue engine that can truly "role-play."

It's not a chatbot — it's a "scriptwriter + actor" — capable of understanding character personality, world-building, and plot direction to generate dynamic dialogue that fits the character's identity.

⚡ Core Technical Highlights

GLM-4.5-Air Foundation

Based on Zhipu GLM-4.5-Air architecture with native Chinese-English bilingual capability
106B parameters (~12B active), inference efficiency far exceeding同等 Dense models
Excellent Chinese comprehension and generation, especially suited for Eastern fantasy题材

Narrative-Aware Fine-Tuning

Character Personality Injection: Dynamically adjusts output style based on character traits (personality, identity, verbal tics, tone words)
Plot Coherence: Supports cross-turn dialogue state tracking — NPCs remember interaction history with players
Context Awareness: Generates differentiated dialogue based on current scene, time, and NPC mood state

MLX 4bit Quantization

MLX framework INT4 quantization, model size compressed to ~60GB
Mac mini M4 Pro 64GB inference speed at ~28 tokens/s
Retains 97%+ of narrative generation quality

Structured Output

Supports JSON-format dialogue node output, seamlessly integrating with game dialogue systems
Built-in tone annotation, action hints, and expression tags for frontend rendering

🎯 Training Data

Data Type	Scale	Description
Classic RPG Dialogue Scripts	200+ Games	The Witcher 3, Baldur's Gate 3, Final Fantasy, Chinese Paladin, GuJian
Original Character Dialogue	8M+ Turns	NPC daily dialogue, quest dialogue, combat dialogue, friendship dialogue
Novels & Screenplays	500K+ Pages	Web fiction, film scripts, stage play dialogue
Game Design Documents	100K+ Pages	Character profiles, plot outlines, world-building bibles
Player Interaction Logs	3M+ Entries	Desensitized real player-NPC interaction data

Training Pipeline

GLM-4.5-Air Foundation
    │
    ├── SFT Phase 1: General Dialogue Alignment (50K high-quality dialogue samples)
    │
    ├── SFT Phase 2: Game Narrative Specialization (character personality + context awareness)
    │
    ├── SFT Phase 3: Sakura Dream Sea Production Data Fine-Tuning (200K real dialogue turns)
    │
    └── RLHF: Narrative Quality Preference Alignment (50K human-annotated dialogue quality ratings)

💻 Hardware Requirements

Configuration	Recommendation
Mac	M2/M3/M4 series, 64GB Unified Memory
PC	RTX 4090 (24GB) + 64GB System RAM
Format	MLX 4bit (INT4)
Speed	~28 tokens/s (Mac mini M4 Pro 64GB)

🚀 Use Cases

Dynamic NPC Dialogue Generation — Generate personalized lines in real-time based on character settings and context
Branching Dialogue Tree Generation — Auto-generate multi-branch dialogue options based on plot nodes
Friendship-Based Dialogue — Generate different tones and content based on player-NPC affinity levels
Random Chatter — Generate海量 non-repetitive daily dialogue for NPCs
Quest Dialogue — Dynamically generate guiding dialogue based on quest progress
Narrative Writing Assistant — Help designers generate plot dialogue and character monologues

🎮 Real-World Validation: Sakura Dream Sea

Sakura Dream Sea (樱梦海) (Steam Store Page) — An Eastern Fantasy Open-World MMO Adventure

The IceBlink model has been fully deployed in Sakura Dream Sea's NPC dialogue system, serving as the core narrative engine powering dynamic in-game interactions:

Live Features

Feature	Description
Dynamic NPC Dialogue	NPCs in the Astral Soul system (e.g., Tami) have every line generated in real-time by IceBlink
Tone Words + Typewriter Effect	Dialogue includes character-specific tone words, paired with typewriter animation
Friendship Dialogue	NPC attitude and dialogue style dynamically change based on friendship level
Context Awareness	The same NPC says different things in different scenes and times
AI Dialogue Interception	Smart decision on whether to trigger AI dialogue or use preset text

Comparison

Dimension	Traditional Preset Dialogue	IceBlink-Driven
Line Count	10-50 per NPC	Theoretically infinite
Repetition	High (same content on repeat interactions)	Extremely low (every conversation varies)
Character Personality	Relies on manual writing	Auto-generated from character settings
Development Cost	High (each line needs writing + VO)	Low (set character, auto-generate)

Technical Architecture

Player Clicks NPC
    │
    ▼
DialogueView (Unity/SLua)
    │
    ├── Check Cached Dialogue
    ├── Build Context (Character + Friendship + Scene + History)
    └── Call IceBlink Model
            │
            ▼
    Generate Dialogue Text + Tone + Action Tags
            │
            ▼
    Typewriter Rendering + Voice Synthesis

📝 Output Format Example

{
  "npc_id": "tami_001",
  "npc_name": "Tami",
  "mood": "cheerful",
  "dialogue": {
    "text": "Oh hey, adventurer! You're back again? Want to hear about what's been happening near the forest today?",
    "tone": "enthusiastic",
    "action": "leans_forward",
    "expression": "smile_warm"
  },
  "options": [
    {"id": "opt_1", "text": "What's going on?", "mood_required": null},
    {"id": "opt_2", "text": "I'm busy today, maybe next time.", "mood_required": null},
    {"id": "opt_3", "text": "(pats Tami's head)", "mood_required": "friendship_5"}
  ],
  "context": {
    "friendship_level": 5,
    "location": "village_square",
    "time_of_day": "morning",
    "quest_progress": "chapter_2_active"
  }
}

🔗 Ecosystem

IceBlink is part of the broader LongShu game AI ecosystem:

Model	Focus	Repository
IceBlink (冰眸)	NPC Dialogue & Narrative	This model
REAP-Architect (天策)	Game Dev Commander & Architecture	GitHub: LongShuGameDev

The REAP-Architect model handles game system architecture, task decomposition, and multi-agent scheduling for game development workflows — while IceBlink powers in-game NPC narrative. Together, they form a complete AI pipeline from development to runtime.

📄 License

Apache 2.0 License

Citation

@misc{longshu-iceblink-2026,
  title={LongShu-IceBlink: A Narrative-Aware LLM for Game NPC Dialogue Generation},
  author={LongShu Team},
  year={2026},
  url={https://huggingface.co/luoyike2003/longshu-iceblink-106b-mlx-v2},
  publisher={HuggingFace}
}

LongShu · IceBlink — Giving Every NPC Its Own Soul

Downloads last month: 373

Safetensors

Model size

107B params

Tensor type

BF16

U32

MLX

Hardware compatibility

4-bit