What do I need to use AI to decompile 3DS, Wii U, and Switch games?

MrRandomNEEDSHELP

New Member
Newbie
Joined
Oct 12, 2026
Messages
1
Reaction score
0
Trophies
0
Age
26
XP
7
Country
United States
Hey everyone! I’m interested in using AI to help with game decompilation for Nintendo 3DS, Wii U, and Switch games, and I’m wondering what tools and setup I would need to get started.

I want to use AI to help analyze game code, identify functions, understand assembly, recreate source code, and hopefully speed up the process of matching the original game's behavior.

A few questions:

  • Which AI tools or coding models are best for decompilation and reverse engineering?
  • Would tools like Claude, ChatGPT, or local coding models be useful for this?
  • What PC specs would I need, especially if I want to run AI models locally?
  • What reverse-engineering tools should I use alongside AI, such as Ghidra or IDA?
  • Which programming languages and assembly architectures should I learn?
  • Can AI realistically help decompile games for all three platforms, or are some much harder than others?
  • Are there existing AI-assisted decompilation projects or workflows I should look into?
  • Would it be better to start with a small game or an existing decompilation project?
I’m interested in learning how to use AI to automate as much of the process as possible while still being able to rebuild the game and verify that it works correctly.

Any recommendations for AI models, tools, tutorials, or beginner-friendly workflows would be appreciated. Thanks!
 
Disclaimer: This was written for informational and entertainment purposes only. Caveat emptor. (Buyer Beware).

1. Most projects use Claude with claude code followed by ChatGPT with codex
2. LocalLLM Redditors reported using Qwen3.8-Flash-Next (125b_6a), GLM-5.3 Flash (321b_18a), and Qwen3.8-27b (27b_27a)
2.1 You will need a local harness as well. Popular choices: OpenCode, Pi, DeepSeekHarness, Hermes
3. Spec Requirements: ~1GB RAM / VRAM per 1B parameters. (Most models are run at Q4 quant or higher)
Memory Bandwidth/GB of active parameters = Tok/s | 256/18 = 14.2 tps
3.1 Common choices: Mac Studio (460-1228GB/s), Mac Mini (307GB/s), DGX Spark (273 GB/s), AMD Strix Halo (256 GB/s)
4. Both ghidramcp and ida-mcp are used
5,6. You will need to learn what tools are required for each platform, read: https://decomp.wiki/
7,8. https://decomp.me/ and https://decomp.dev/ have existing projects and individual functions to decompile.
8.1 The bigger the function the harder it is to decompile. A small game might have big functions.

Edit: Additional websites: https://recomp.fyi/ https://bytematch.dev/
 
Last edited by radicalten,

Site & Scene News

New Hot Discussed User Submitted