---
language:
  - zh
  - en
license: apache-2.0
tags:
  - moe
  - code-assistant
  - gamedev-assistant
  - llm
  - quantized
  - mlx
  - text-generation
  - gamedev
  - roleplay
  - multi-agent
  - game-development
  - unreal-engine
  - unity
  - vision
  - multimodal
pipeline_tag: text-generation
library_name: mlx-lm
base_model:
  - 0xSero/Qwen3.5-99B
base_model_relation: quantized
---
# LongShu · Reaper-V1 (99B FP16 Edition)
An ultra-large multimodal MoE model with 99B total parameters / 10B active, featuring 205 experts and native vision support. Quantized with FP16 baseline for maximum precision on consumer-grade Apple Silicon hardware.

github: https://github.com/luoyike2003ls/LongShuGameDev

---

## 🎯 Model Overview

This is the **FP16 baseline** variant of the LongShu Reaper series, built on the massive Qwen3.5-99B MoE architecture with **205 routed experts**. Unlike the standard 122B-A10B version, this model features:

- **Significantly larger expert pool** (205 vs 105 experts) for richer knowledge capacity
- **FP16 quantization baseline** (not BF16) for better compatibility and precision
- **Native multimodal support** with vision config injected for image understanding
- **3.585 bits-per-weight** extreme compression retaining maximum reasoning capability

---

## 🏗️ 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 |
| 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 |
| Coordinator | Hengshu | Team collaboration expert | Intelligent task allocation, cross-functional coordination |

---

## ⚡ Core Technical Highlights

### Massive MoE Architecture
- **48-layer** deep network with hybrid Linear + Full Attention
- **205 routed experts**, only a subset activated per token
- Enormous knowledge capacity with efficient sparse computation

### FP16 Precision Baseline
- Quantization coefficients stored in **float16** (not bfloat16)
- Better compatibility across platforms and inference engines
- Embeddings protected at **6-bit** for maximum semantic fidelity

### Game Engine-Aware Hybrid Quantization Strategy
| Component | Bits | Group Size | Rationale |
|-----------|------|------------|-----------|
| Embeddings & LM Head | 6-bit | 64 | Semantic fidelity hub |
| Attention Layers | 4-bit | 64 | Core reasoning mechanism |
| Shared Experts | 4-bit | 64 | Critical path components |
| Routed Experts | 3-bit | 64 | Massive pool, extreme compression |
| Router Gate | 4-bit | 64 | Decision-critical routing |

### Vision Multimodal Support
- Vision config injected from original model
- Supports `image_token_id`, `video_token_id`, `vision_start/end_token_id`
- Ready for visual understanding tasks in game development

---

## 📊 Model Specifications

| Metric | Value |
|--------|-------|
| Total Parameters | ~99B |
| Active Parameters | ~10B |
| Number of Experts | 205 |
| Expert Layers | 48 |
| Quantization | 3.585 bpw (FP16 baseline) |
| Model Size | ~41 GB |
| Context Length | 262,144 tokens (256K) |
| Inference Speed | ~35 tokens/s (Mac mini M4 Pro) |
| TTFT (First Token) | ~10.75s |

---

## 💻 Hardware Requirements

| Configuration | Recommendation |
|---------------|----------------|
| **Mac** | M2/M3/M4 series, 64GB+ Unified Memory |
| **PC** | Dual RTX 3090/4090 (24GB+) |
| **Format** | MLX 4-bit with FP16 baseline |
| **Speed** | ~35 tokens/s (Mac mini M4 Pro 64GB) |

---

## 🚀 Quick Start

```python
from mlx_lm import load, generate

model_path = "luoyike2003/LongShu-Qwen3.5-99B-Reap-GameDev-MLX-4bit-FP16"
model, tokenizer = load(model_path)

prompt = "Explain the architecture of a scalable MMO server system"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=text, max_tokens=2048)
print(response)
```

---

## 🎮 Use Cases

- **Game System Architecture Design** — Decompose complex requirements into modular architectures
- **Unreal/Unity Engine Expert** — Deep understanding of engine internals and APIs
- **AI Behavior Tree Design** — Design and optimize game AI systems
- **Code Review & Optimization** — Review code quality, identify performance bottlenecks
- **Multimodal Game Asset Analysis** — Analyze game screenshots, UI designs, and visual assets

---

## 📄 License

Apache 2.0 License

---

*LongShu · AI-Powered Partner for Game Development*

<!-- Updated: 2026-06-19 -->
<!-- model: 99B-REAP-FP16 -->