Analyzing the News: Meta's Muse Spark Open

Thirteen outlets covered Meta's new open-weight AI model. Fíor scored twelve of them and read only one as AI.

Fíor is a natural language classifier that predicts whether a piece of text was more likely written by a person or by AI.

Meta released Muse Spark Open on 15 October, with weights that anyone can download. Meta says the model has 816 billion parameters, 42 billion of them active at a time, and can take in up to a million tokens at once. It is built for coding and for agents that carry out long, multi-step tasks.

Meta worked with Apple to make it run on two M5 Ultra Mac Studios, using Apple's MLX framework. Meta calls it the most capable US open-weight model to date. That is the company's own claim. The licence allows commercial use on limited terms.

The outlets were Reuters, AFP, Bloomberg, CNBC, Axios, the Financial Times, TechCrunch, The Verge, VentureBeat, MacRumors, 9to5Mac, TechRepublic and SiliconANGLE. That is two wire services, the business press, and seven technology outlets. The AFP piece was scanned as published by the Guardian. Fíor picked up the Financial Times paywall rather than the article, so that result is left out.

The result

Eleven read as human: Reuters, AFP, Bloomberg, CNBC, TechCrunch, The Verge, VentureBeat, MacRumors, 9to5Mac, TechRepublic and SiliconANGLE.

One read as AI: Axios.

That is a long way from the last two pieces in the series. The IMF forecast and US inflation both split close to evenly.

The samples are not like for like. Seven of the twelve here are technology outlets, and none of them appeared in the earlier pieces. Of the outlets that did, Reuters and CNBC read as human in all three. AFP moved the other way: it read as AI on the IMF forecast and as human here.

The one that read as AI

Axios writes in a fixed house format. Its pieces are built from short blocks under set headings, such as "Why it matters" and "The big picture". This one also carried original reporting, including an interview with Meta's head of AI and detail on how the two companies came together.

A fixed format is the obvious explanation. It does not hold up on its own. Reuters has a strict house style too, and it has read as human in all three pieces so far.

Why launch coverage reads like this

Every outlet started from the same source: Meta's announcement, with the same headline numbers. Most repeated them.

What they did next varied far more than in the earlier pieces. CNBC, The Verge and Axios went to the Apple partnership. MacRumors and 9to5Mac went to the hardware. Bloomberg looked at why Meta is giving this model away while keeping its best one closed. AFP and VentureBeat compared it with Chinese open models. SiliconANGLE went through the licence and the safety testing.

A data release leaves little room to choose an angle. A product launch leaves a lot. That may be part of why this coverage reads so differently.

Shared numbers alone do not explain much. Nearly every piece carried the 816 billion figure, and eleven still read as human.

Axios may have had AI help, with a first draft or a final polish. Or its format is fixed enough that a reporter and a model would produce the same shape. Fíor cannot tell those apart on its own.

Meta's pitch is that a model this capable now runs on two desktop computers. Once that is true, text that reads as AI does not have to come from a few companies' servers. It can come from a desk anywhere.

What Fíor is

Fíor is a machine learning classifier, currently in beta and available as a browser extension. It gives a prediction, not a verdict. It cannot see who sat at the keyboard, and nothing here is a claim about any journalist or publication. Reading as AI and being written by AI are not the same thing.

Treat it as a signal, not proof, and check the pages you read yourself.

Try the browser extension at Fíor.

Analyzing the News is a series where we run one major story through Fíor and see what comes back.