01What Is Muse Spark 1.2 and How It Works

Muse Spark 1.2 is Meta's latest AI model for building interactive experiences like games. In this episode, Julian Goldie tested it by creating various game builds and comparing the results with other models.

Muse Spark 1.2 works by generating playable game environments, characters, and controls from prompts. It also includes Muse Code, an agentic harness that can be integrated into agent systems to perform tasks like learning from guides and creating skills. The model aims to be a versatile tool for both direct coding and agentic use.

02Main Features and What's New in Muse Spark 1.2

Muse Spark 1.2 includes several features that Julian highlighted during testing.

  • Game generation: It can create a variety of game types, including RPG style, flight simulators, racing games, and Doom-style shooters.
  • Muse Code agentic harness: This allows the model to be used for agentic tasks, such as analyzing guides and creating skills, when integrated into systems like Hermes.
  • Speed in agentic tasks: In a comparison, Muse Spark 1.2 replied faster than Quen 3.8 on a learning task.
  • Integration capability: It can be plugged into agent operating systems, enabling delegation as a sub-agent.

Julian noted that while these features are present, performance in game builds lagged behind other models.

03How to Set Up and Get Started with Muse Spark 1.2

To start using Muse Spark 1.2, you need to integrate it into an agent operating system or use it directly with a coding environment. Here are the steps Julian showed:

  1. Access an agent operating system like the one from AI Profit Boardroom that supports Muse Code.
  2. Select Muse Spark 1.2 as your model in the system's settings.
  3. For agentic tasks, use a platform like Hermes agent and choose Muse as the model.
  4. Provide a task, such as a learning command, and observe its response.
  5. For coding, input prompts to generate game builds directly and preview them.

Julian emphasized that setting up is straightforward but results vary by task type.

04How to Use It: Examples from the Episode

Julian tested Muse Spark 1.2 on multiple game builds and an agentic task.

For games, he prompted it to create an RPG style game, a flight simulator, a Neon Blaster game, Twilight Vale, a racing game, Dragon Realm, a crypt game, and a parachute game. Each build showed issues like lag, bugs, poor controls, and lack of detail. For example, Twilight Vale spun around when moving backwards, and the crypt game had missing arms on characters.

For agentic use, he used a learning task in Hermes where Muse Spark 1.2 had to analyze a guide and create a skill. It replied faster than Quen 3.8 and produced a detailed skill with specific information and tables.

He concluded that using Muse Spark 1.2 as a sub-agent might be a better use than direct coding.

05Results and Comparisons Shown in the Episode

Julian compared Muse Spark 1.2 against GPT 5.6 Sol, Fable 5, and Quen 3.8 on game builds. Consistently, GPT 5.6 Sol and Quen 3.8 produced better quality games with richer detail, better controls, and fewer bugs.

For example, Fable 5's version of a dragon flight game had impressive lighting and shadows, while GPT 5.6 Sol created more interesting environments. Quen 3.8 was praised for its output despite some users saying it wasn't great. Muse Spark 1.2 never won on any task.

However, in the agentic test, Muse Spark 1.2 replied faster than Quen 3.8 and created a solid skill. Julian noted that Quen 3.8 was slow in agentic tasks, while Muse Spark 1.2 performed better in that context.

06Tips, Limits, and Mistakes to Avoid

Julian shared several observations about using Muse Spark 1.2.

  • Avoid using it for complex game builds directly; it lacks the quality of frontier models.
  • Expect lag and bugs in generated content, especially in interactive environments.
  • For agentic tasks, it can be useful as a sub-agent or within systems like Hermes.
  • Don't expect it to replace frontier models like GPT 5.6 Sol or Quen 3.8.
  • Use tools like Goldie Bench to test models in realistic scenarios, not just benchmarks.

He noted that Meta models have never impressed him, and Muse Spark 1.2 continues that trend for direct coding.

07Who It Is For and Why It Matters

Muse Spark 1.2 may be suitable for users who want a model that can handle agentic tasks and can be integrated into agent operating systems. It offers speed in certain agent-based operations, as shown in the learning task.

However, for developers building games or requiring high-quality generated content, this model is not ideal. Julian suggests that large organizations might use it as a sub-agent for delegating simple tasks, but for serious production, frontier models are better.

The release matters because it shows Meta's push into agentic AI, even if the model's direct coding capabilities lag behind competitors.