I Ran Meta Spark Open Locally for 24 Hours... USA's #1 Local AI 🫣

Muse Spark Open Edition: Local Model Review

The video tests Muse Spark Open Edition, a local, quantized version of Meta’s Muse Spark model, mainly by asking it to generate code and browser-based demos. The presenter describes it as a strong US open-weight model, but finds its performance uneven across tasks.

Benchmarks and local setup

• The presenter cites a score of about 60 on SWE-bench for this version, below the roughly 75.4 cited for Muse Spark 1.3. He characterizes it as competitive with some other open models, while cautioning that benchmark strength did not consistently translate into his coding demos.
• The model’s chat template exposes reasoning levels—low, medium, high, and extra high—as numeric settings. The presenter notes that these values can potentially be changed, though he does not systematically test custom settings.
• The tested quantized setup used roughly 435–500 GB of memory, making it suited to high-memory workstations. He reports that a standard Q4 quant often produced broken code, while his modified “INF” quantization ran more reliably. He says he uploaded a quantized version.

Coding and generation tests

The presenter’s most extensive tests ask the model to build interactive web projects, often in HTML and JavaScript.

3D scenes and simulations

The model produced a clickable solar system, a volcanic eruption with lava damaging a town, and a basic 3D face. Some results looked convincing for prototypes; the solar system was relatively simple, and the face lacked some features.

Games

The model generated demos inspired by Call of Duty, GTA, Tomb Raider, Red Dead Redemption, Super Mario, Final Fantasy 7, and other games. Results ranged from playable but rudimentary to amusingly flawed. A go-kart track was among the more successful examples; some games had weak controls, incomplete mechanics, or odd visuals.

Image-to-code

Given screenshots, the model recognized an Age of Empires-style scene and recreated a simple version of another mobile game. The results captured some of the source game’s structure, but were not necessarily fully playable.

Python

A solar-system prompt yielded multiple implementations, including a 3D plotting version and a browser-based Plotly version. The presenter found these more reliable than standard Q4 outputs.

Productivity-app clones

The model generated basic Word, PowerPoint, Photoshop-like, and iMovie-like interfaces. Some features worked—such as editing PowerPoint text, duplicating a slide, and manipulating a layer—but the apps were incomplete or had bugs.

Other coding

A weather-widget animation was praised as visually polished. A Logic Pro-style music tool could generate and play simple melodies, though it also produced a runtime error.

The presenter says the model often offered suggestions for follow-up improvements rather than building a fully featured result immediately. He also reports three runtime errors across about 30 generations, while noting that some tasks—particularly shader work for a planet demo—stalled or failed. Overall, he sees it as capable of useful prototypes, but not consistently at the level implied by its benchmark standing.

Reasoning, coding knowledge, and modalities

• In several logic and math checks, the model gave answers the presenter judged correct, including an integer problem and an International Mathematical Olympiad-style question.
• It also handled examples involving Swift, C++, and a Windows Runtime API, though one API answer was described as partly correct and partly incorrect.
• Audio results were inconsistent: the presenter initially describes the model as supporting audio inference, but his tests suggest it often treated audio as text, became confused when asked to transcribe, or said it could not analyze audio. He did not demonstrate video inference.
• A quantized version unexpectedly produced more provocative role-play responses in one test. The presenter says this happened while testing audio and does not show the more explicit outputs.

Overall assessment

The presenter is positive about Meta returning to open-weight models and considers Muse Spark promising, especially for local prototyping. His review is qualified: quantization and memory requirements matter, and generated code can range from useful and polished to incomplete or broken. He also notes that he did not have time to show every improved-quantization result.

Guides and demonstrations

The video functions as a hands-on review rather than a step-by-step installation tutorial. Demonstrations include:

• Comparing standard Q4 with the presenter’s modified INF quantization.
• Adjusting reasoning settings through the chat template.
• Testing code generation, image-based prompts, logic, math, and audio.
• Showing how follow-up prompts could add features to generated projects.

Main speaker/source: The xCreate channel presenter, providing commentary and reporting results from his local tests. No other speaker is identified.