LongShu ยท REAP-Architect

TL;DR: A multi-agent collaboration core brain specifically crafted for game development, based on Qwen3.5-122B-A10B-MoE architecture with deep optimization, running smoothly on consumer-grade hardware.


๐ŸŽฎ Model Positioning

Traditional general-purpose LLMs often suffer from severe "hallucinations" or provide superficial suggestions when facing complex system architectures, deep engine APIs (Unreal/Unity), and state machine logic in real game development scenarios.

LongShu aims to solve this pain point. Based on the powerful Qwen3.5-122B-A10B-MoE foundation, we've performed deep "surgical" domain-specific fine-tuning to create the core brain of this Multi-Agent System โ€” the Commander Model (TIANCE).

It doesn't just understand code; it understands game engineering. It's no longer a chatbotโ€”it's a true "Technical Director + Lead Architect".


๐Ÿ—๏ธ REAP Ecosystem

LongShu is the central brain of a complete game development agent network:

Role Codename Positioning Core Capabilities
Commander Tiance Core brain, logic reasoning hub Global planning, system decomposition, task dispatch
Architect Xuangou Code architecture expert Tech structure analysis, architecture optimization, tech debt identification
Executor Moxing Task execution specialist Coding, debugging, test case generation
Watcher Zhuzhao Monitoring & alerting expert Log analysis, anomaly detection, risk early warning
Scholar Wenyuan Knowledge management expert Documentation understanding, knowledge graphs, intelligent retrieval
Coordinator Hengshu Team collaboration expert Intelligent task allocation, cross-functional coordination

โšก Core Technical Highlights

Hybrid Attention Architecture

  • 60-layer network with 3:1 alternating Linear + Full Attention
  • Balances lightning-fast linear inference with full attention precision

Long Context Support (256K)

  • Native support for 262,144 token context
  • Can ingest entire project headers, design docs, and API documentation in one shot

Extreme MoE Sparsity

  • 105 experts, only 10 activated per token
  • High compute efficiency, inference speed of ~35 tokens/s (Mac mini M4 Pro 64GB)

Game Engine-Aware Hybrid Quantization

  • Core 20 experts: high-precision IQ4_NL/Q5_K
  • Non-core 85 experts: extreme compression IQ2_XXS
  • Massive model size reduction while retaining 98.5% core reasoning capability

๐ŸŽฏ Specialized Training Data

Data Type Scale Description
Real Game Projects 50+ MMO, FPS, ARPG, Roguelike genres
Core Source Code 2B+ Tokens UE (C++), Unity (C#), Godot, Lua hot-reload frameworks
Engineering Docs 300K+ Pages GDDs, system breakdowns, game design logic, performance analysis
High-Quality Online Data 10B+ StackOverflow gamedev, GitHub Issues, graphics papers

๐Ÿ’ป Hardware Requirements

Configuration Recommendation
Mac M2/M3/M4 series, 64GB Unified Memory
PC Dual RTX 3090/4090 (24GB)
Format llama.cpp IQ3S extreme compression
Speed ~35 tokens/s (Mac mini M4 Pro)

๐Ÿš€ Use Cases

  • Automated Test Case Generation - Auto-generate test plans based on code logic
  • Daily Build Error Diagnosis - Analyze compile/runtime errors with fix suggestions
  • Level Toolchain Dispatch - Understand design requirements, dispatch executors
  • System Architecture Design - Decompose complex requirements into modular architectures
  • Code Review - Review code quality, identify potential issues

๐ŸŽฎ Real-World Application

Sakura Dream Sea (ๆจฑๆขฆๆตท) (Steam Page) โ€” An Eastern Fantasy Open-World MMO Adventure

The Sakura Dream Sea development team serves as one of the core pilot users of the LongShu model, deeply integrating the complete LongShu agent capabilities into their game development pipeline:

  • Leveraging the Tiance/Commander Model for global task planning and module decomposition
  • Utilizing Xuangou/Architect for game system architecture optimization and code review
  • Accelerating core gameplay development (combat systems, AI behavior trees) through Moxing/Executor
  • Monitoring server performance and anomaly logs with Zhuzhao/Watcher

This open-world MMO set in the "Eternal Sakura Continent" represents the best practice validation of LongShu large models in game industrial production. As you embark on this fantasy continent adorned with falling cherry blossoms, adventuring alongside those "breathing AI companions," it's the LongShu agent network powering the intelligent operation of the entire game world.


๐Ÿ“„ License

Apache 2.0 License


LongShu ยท Your Intelligent Partner for Game Development

Downloads last month
418
Safetensors
Model size
97B params
Tensor type
BF16
ยท
U32
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support