THE EXPERIMENT BEHIND THE ARCADE
Local models.
Real challenges.
Playable results.
Gencade showcases agentic coding challenges: a way to compare how self-hosted models plan, write code, solve problems, and turn a game brief into something you can actually play.
Challenge runs use locally hosted models, with no hosted inference: no Claude, Codex, or other cloud-hosted frontier models.
What are we testing?
The games are the results of coding challenges designed to explore agentic capabilities. How well can a model understand a brief, make implementation decisions, use tools, debug its own work, and deliver a complete game?
Trying different locally hosted models lets us compare their coding abilities through concrete, playable outcomes. The arcade is a place to explore those results, including their rough edges, rather than a claim that every experiment is a finished product.
Why games?
A game asks a coding agent to bring many things together: controls, simulation, visuals, performance, and a working gameplay loop. It gives us something tangible to inspect and a simple question to ask: does it work, and is it fun?
Pick a game to launch it in your browser. Device icons show whether it supports a PC, controller, or phone. Support varies by game; phone support means touch controls are implemented.
THE CURRENT TEST PLATFORM
Consumer hardware. Self-hosted inference.
Challenge models run on the FreeToken server. Here’s the current hardware and the resources allocated to its inference environment.
- PROCESSOR
- AMD Ryzen 7 3700X 8 cores / 16 threads on the host
- GRAPHICS
- NVIDIA RTX 3070 Ti 8 GiB of VRAM
- HOST MEMORY
- 64 GB RAM 56 GiB allocated to FreeToken
- ENVIRONMENT
- Ubuntu on Proxmox LXC container / 12 allocated CPU cores
Hardware and allocation checked 3 November 2026. Model choice and configuration can vary between challenge runs.