In a recent experiment by Atomic Chat, Meta’s Muse Spark 1.1 successfully created three self-playing 3D arcade games at a significantly lower cost compared to Google’s Gemini 3.6. Muse Spark 1.1 utilized 31.3K tokens to produce these games for just $0.026, while Gemini 3.6, which used 28.8K tokens, cost $0.073 and struggled with rendering details and smooth gameplay. This comparison showcases the advantages of Atomic Chat’s offline LLM environment, which enables local execution without cloud dependence, as well as the differences in handling complex physics simulations and game logic between the two AI models.

Meta: Meta is the technology company that develops and releases advanced AI models under its research initiatives. The company is directly tied to the news through its Muse Spark 1.1 model, which was tested head-to-head in the experiment shared by Atomic Chat.
Gemini 3.6: Gemini 3.6 is Google’s AI model that was compared against Muse Spark 1.1 in the same controlled test. It processed a similar token count but produced lower-quality outputs with glitches, incomplete backgrounds, and higher overall cost for the identical set of 3D self-playing game tasks.
Atomic Chat: Atomic Chat is an open source desktop application that runs large language models entirely offline on users’ personal computers as a private alternative to cloud-based services like ChatGPT. The project, associated with @atomic_chat_hq, published a GitHub repository and download links for the tool. In this news, Atomic Chat conducted and shared the results of a controlled experiment pitting two AI models against each other on a specialized 3D game generation task.
Muse Spark 1.1: Muse Spark 1.1 is an AI model developed by Meta and evaluated in the Atomic Chat experiment for its ability to generate self-running 3D arcade and tabletop games. It completed the full set of three physics-based game prompts with higher visual quality and detail than the competing model while incurring lower cost.
@atomic_chat_hq: @atomic_chat_hq is the X account and project handle for Atomic Chat, the open-source offline LLM desktop app. It originated and posted the full experiment details, including prompts, outputs, token counts, and cost comparisons between the two models.

Local AI Execution: Atomic Chat provides a fully offline desktop environment for running LLMs without reliance on cloud services.
AI Model Comparison: The experiment highlighted differences in how Meta and Google models handle complex 3D physics simulation and self-contained game logic.