Research Methodology

1. Meta's own published material — the Muse Code announcement on research.meta.ai, the Muse Spark posts on ai.meta.com, and the developer documentation. Treated as authoritative for what the product claims to do.
2. First-hand artefact evidence — the contents of the installer and launcher scripts Meta serves publicly. These were read as text and never executed. They are the strongest available evidence for install paths, supported platforms, the authentication flow and error strings, because they are the shipping code rather than a description of it.
3. Independent third-party measurement — evaluations run by parties with no stake in the result, such as Artificial Analysis. Reported separately from Meta's own figures, never averaged with them.
4. Reputable press — used mainly to relay figures from Meta documents we could not read directly, and always identified as second-hand.
5. Community reports — Hacker News, X and developer forums. Useful for symptoms, workarounds and sentiment. Never treated as fact on its own.

Where a claim rests only on community reporting, the page says so in the sentence itself — "community reports suggest", "users report", or an explicit attribution to the person who made the claim — and links the post. Where a price or specification comes from third-party trackers because Meta's own pricing page renders through JavaScript and cannot be quoted directly, the page marks it "as reported" and links Meta's documentation so you can check the current figure yourself.

Benchmark figures are the most common place this goes wrong elsewhere, so they get the strictest treatment. Meta's self-reported scores and independent evaluations are kept in separate labelled rows, because the two use different harnesses, different attempt counts and different comparison sets. A page that presents them as one number is telling you something that was never measured.

When something is simply unknown — Muse Spark's parameter count, its knowledge cutoff, whether Windows support is planned — the page says it is unpublished rather than filling the gap with an estimate.

We do not run our own benchmarks, so no score on this site originates here. We do not publish a hands-on review of Muse Code, because writing one would require sustained real-world use and inventing that experience would be worse than not having it. We do not reproduce extracted system prompts or other material lifted from the binary; where the community has reported such findings, we summarise what was claimed and link the report.

We do not host the installer, mirror any binary, or publish a download link of our own. Installation instructions quote Meta's command and link Meta's documentation.

Every page carries a visible last-updated date at the top and the same date in its structured data. Pricing, rate limits and model specification pages are re-checked against Meta's documentation weekly during the beta, and any page whose subject changes is updated the same week, with the change reflected in the site's sitemap so search engines see the revision.

The full sourcing list for the underlying research sits in the repository behind this site, and each page repeats the sources it actually relied on at the bottom of that page.