Show HN: Exploring the limits of local LLM coding on consumer HW (github.com)

Infinity Arcade is an open-source Show HN project that uses local LLMs (via a Lemonade Server) to generate playable retro-style games on the fly: you type a prompt and the system returns a pure-Python pygame game you can run immediately. It’s available via pip or from source, caches generated games under ~/.infinity-arcade/games, and requires Lemonade Server running at http://localhost:8000 with models installed. Generated games follow strict constraints (only pygame, no external assets, proper event loops) and the repo includes example prompts, troubleshooting tips, and notes that outputs can be imperfect and may need regeneration or manual fixes. For the AI/ML community this is a compact, practical exploration of local LLM coding and the hardware tradeoffs for on-device code generation. The project auto-detects hardware and recommends models — minimum: Ryzen AI 7000+ CPU with 32GB RAM and a small Playable1-GGUF model; suggested setups include iGPU Ryzen AI 300-series, dGPU Radeon 7800XT (16GB VRAM) or Strix Halo APU with larger unified memory using Qwen3-Coder-30B. It highlights real constraints (memory, VRAM, disk), shows how consumer GPUs/APUs can be used for instant code synthesis and prototyping, and serves as a reproducible testbed for evaluating model/code-quality, latency, and reliability when running LLMs locally.