The problem: while LLM coding is amazing with large cloud-based models, the results you get with today's open-source 7B-8B parameter models (the maximum size for an average consumer laptop) are anything but inspiring. Even when we constrain ourselves to something "easy", like vibe coding retro arcade games in Python, the best sub-8B-parameter models can barely produce working copies of Snake and Pong.
Our solution: first, we needed an app that would make it easy to drive local LLMs with exactly the right prompts and features for creating games. We built three minimal agents: Create, Remix, and Debug. This allows users to make new games, make changes to existing games, and allows the system to automatically fix its own bugs.
Second, we needed to build our own model because none of the open-source offerings were cutting it. We assembled a dataset with >50k lines of high-quality Python game code, then used LORA SFT to create Playable1-GGUF (https://huggingface.co/playable/Playable1-GGUF): by far the world's highest performing 7B model at this task. It can handle 3x more game types and many fun riffs, like "space invaders with exploding bullets" and "breakout where the ball accelerates every time, I hit it."
There's a short GIF with a demo of the outputs here: https://github.com/lemonade-sdk/assets/blob/main/arcade/arac... And one showing the full UI here: https://github.com/lemonade-sdk/assets/blob/main/arcade/infi...
Ultimately, the goal is to deliver a reference design that inspires startups to build with local LLMs. Local has no ongoing cloud costs and no privacy/security concern with sharing user's data over the internet. The main challenges are the capabilities of smaller local models, as well as the work to integrate a software stack that is relatively less mature than OpenAI/Open Router APIs. We believe Infinity Arcade shows that both challenges can be overcome and the benefits of local can be realized.
What you can do today: download the app from GitHub, install with 1 click, run Infinity Arcade and its LLM 100% free and local, check out the training data and process for the fine-tuned model, and build your own app using all of this as a reference.
Links:
GitHub: https://github.com/lemonade-sdk/infinity-arcade
Model: https://huggingface.co/playable/Playable1-GGUF
We’d love to hear your feedback! Are you convinced that the benefits of local outweigh the challenges? Do you have all the resources you need to get started? Let us know in the comments!