A pixel art space shooter built entirely by a 9B parameter AI model running locally on a single RTX 3060.

One prompt. One GPU. Zero hand-written code.

• Model: Qwen 3.5 9B Q4_K_M
• Hardware: RTX 3060 12GB
• Agent: Hermes Agent by NousResearch
• Inference: llama.cpp at 35-50 tok/s
• Context: 128K tokens
• Iterations: 4 phases, ~6 prompts total
• Lines of code: 3,263 across 13 files

• Pixel art octopus enemies (4 types + boss)
• Particle effects and screen shake
• Procedural audio (no audio files)
• 4-layer parallax background
• Combo system and score tracking
• Level progression with wave clearing
• Boss fights at level 5+
• Power-up drops (health + weapon boost)
• Upgrade tier system
• Full game loop: menu > play > pause > game over > restart

cd space-shooter
python3 -m http.server 8000

Open localhost:8000 in your browser.

• Mouse - Move ship
• Click - Shoot
• ESC - Pause/Resume

RTX 3060 12GB ($250 used on ebay)
  └── llama.cpp (CUDA, flash attention, q4_0 KV cache)
      └── Qwen 3.5 9B Q4_K_M (5.3GB GGUF)
          └── Hermes Agent (31 tools, 11 model-specific parsers)
              └── This game

See prompts/octopus_invaders_prompt.md for the exact prompt used to generate the initial game.

See prompts/iteration3_plan.md for all iteration prompts and tracking.

MIT

@sudoingX using Hermes Agent by @NousResearch