AI Dodge 🎯

A Phaser.js 3.60 dodge game — built by 13 different LLMs from the exact same prompt, ranked and compared.

┌──────────────────────────────────────────────┐
│                                              │
│    🔴                    🔴                  │
│              🔵                               │
│    🔴                    🔴         🔴       │
│                                              │
│         🔴                                   │
│                                              │
└──────────────────────────────────────────────┘
     Dodge the red hunters. Survive. Compete.

🏆 Results — 13 Models

Each model received the same prompt: create a Phaser.js 3.60 dodge game with WASD movement, red chasing enemies, score tracking, wrap-around edges, game over, and R restart.

Model	Provider	Score	Size	Time
Mistral Small ✅	Mistral API	95	6.3 KB	12s
DeepSeek V4 Flash (0.7) ✅	DeepSeek Chat	90	15 KB	~14s
owl-alpha ✅	OpenRouter	88	7.9 KB	53s
DeepSeek V4 Flash (0.4) ✅	DeepSeek API	82	7.1 KB	14s
GPT-OSS-20B ✅	Local (Mac Mini)	78	13.5 KB	23s
Nemotron 3 Ultra ✅	OpenRouter	76	5.1 KB	86s
Qwen 3.5 9B ✅	Local (Mac Mini)	74	4.6 KB	190s
Llama 3.1 8B ✅	Local (Mac Mini)	72	6.7 KB	55s
Gemma-4-12b-qat ✅	Local (Mac Mini)	70	5.2 KB	281s
Gemma-4-12b-coder-fable ✅	Local (Mac Mini)	68	3.2 KB	112s
DeepSeek V4 Pro ⚠️	DeepSeek API	65	6.8 KB	101s
Lfm 2.5 8B ❌	Local (Mac Mini)	—	25.7 KB	128s

Full Blog Post →

🎮 Play Online

https://aigamingdev.com/games/ai-dodge/

• WASD / Arrow keys — Move your blue circle
• Avoid red enemies — They chase you with seek AI
• Survive — Score increases every second
• R — Restart after game over

📄 Game Files

File	Model	Source
index.html	Main game (reference build)	—
ai-dodge-deepseek-v4-flash.html	DeepSeek V4 Flash (interactive, 5 prompts)	Play
deepseek-v4-flash-api.html	DeepSeek V4 Flash (API, 0.7 temp)	Play
deepseek-v4-pro.html	DeepSeek V4 Pro	Play
gpt-oss-20b.html	openai/gpt-oss-20b	Play
llm-deepseek-v4-flash.html	DeepSeek V4 Flash (0.4 temp)	Play
llm-mistral-small.html	Mistral Small	Play
llm-nemotron-3-ultra.html	Nemotron 3 Ultra	Play
llm-owl-alpha.html	owl-alpha	Play
llm-llama-3.1-8b.html	Llama 3.1 8B	Play
llm-gemma-4-12b-coder-fable.html	Gemma-4-12b-coder-fable	Play
llm-gemma-4-12b-qat.html	Gemma-4-12b-qat	Play
llm-qwen-3.5-9b.html	Qwen 3.5 9B	Play
llm-lfm-2.5-8b.html	Lfm 2.5 8B (empty canvas)	Play

🧪 Methodology

All variants tested with Playwright headless Chromium + SwiftShader WebGL. Verified: canvas renders, game loop runs, keyboard input works, no JS errors.

Full 3D benchmark also available: see the 3D Space Dodge repo for 18 models building a Three.js dodge game.

License

MIT