Meta Releases Muse Spark Open, Runs It on Apple Mac Studios

Meta released Muse Spark Open today, calling it "the most capable US open-weight model to date." As proof of what it can do, Meta says it worked with Apple to run the model locally across two or four connected Mac Studios powered by M5 Ultra chips, using Apple's MLX framework.

That is a lot of superlative for one afternoon, so let us read it slowly.

What Meta says it built

According to Meta's announcement, Muse Spark Open is built for coding and agentic work, and it "excels at long-horizon tasks, complex software engineering, and multi-step reasoning." The company lists these specs:

• 816B total parameters, 42B active
• A 1M token context window
• Natively multimodal across text, image, video and audio

Muse Spark Open is similar to Muse Spark 1.3. That is a useful anchor, since it places the release in an existing family rather than presenting it as something from nowhere.

Now the footnote. "Excels" is a verb, not a measurement. The specs above describe size and context length, not performance, and the announcement text gives no benchmark behind "most capable." A comparative claim needs a comparison. Parameter counts tell you how big the model is. They do not tell you how good it is.

The word "open" is doing some work

Bloomberg Tech ran this headline on the same news: "Meta Releases Open AI Model While Keeping Its Best Tech Closed." The headline does not identify which closed technology it means.

The two framings are less opposed than they look. Meta's claim is scoped: most capable among US open-weight models. Bloomberg's is about what sits outside that category. A company can hold the open-weight crown and still keep something better in the vault. Both statements can be true at once, which is the most interesting thing about them. The superlative has a qualifier, and the qualifier is carrying the sentence.

What we also do not have is the license. The terms under which Muse Spark Open is distributed are not in the announcement text. The open-weight label does not resolve the unanswered question of what the license permits.

Two rivals, one rack of Macs

The Apple partnership is the part with a plot. Meta says the demonstration runs the model across two or four connected Mac Studios with M5 Ultra chips, on Apple's MLX framework. It is offered as evidence of capability, though running locally is a deployment fact, not a quality score.

What is missing here is the bill. No price or exact hardware configuration for the connected Mac Studio setup is in the announcement, and that matters for anyone wondering who this is for.

The backstory explains why the pairing draws attention. CNBC reported, as relayed by Techmeme, that "Tim Cook and Mark Zuckerberg have clashed for years, but with John Ternus taking the reins at Apple, the rivals are finding common ground in artificial intelligence."

Meta's AI chief, Alexandr Wang, supplied the family-therapy framing. "We're excited for mommy and daddy to be playing nice again," he told Axios.

What to watch

The release lands in a busy week for open models. Nvidia released updated versions of its Nemotron family on Tuesday, and Reflection AI and France's Mistral both released new open source models last week. Muse Spark Open will be judged against that field, and Meta's claim of the top US spot is now an invitation to check.

Three things would settle the matter:

1. Independent evaluations. Scores on coding and agentic tasks, since that is what the model is built for, run by someone other than the company that wants the headline.
2. The license. Terms decide whether "open-weight" means open for builders or open for press releases.
3. The hardware math. What a two-Mac or four-Mac setup costs, and how the model actually performs on it.

Until those arrive, Muse Spark Open is a large model with a confident announcement and a short list of numbers. The announcement gives us parameter counts and context length, but no benchmark behind the quality claim. Show me the eval, not the vibe.