Meta Keeps Delaying Muse Spark, Its Answer to the Frontier Labs

Two months ago, Meta's chief AI officer Alexandr Wang told developers that Muse Spark would reach them soon. As of this week, they are still waiting. The Wall Street Journal reported that Meta has pushed back the developer release of its new flagship model again, with no firm date on the calendar.

Muse Spark is a notable bet for the company. It is Meta's first closed-source model, built inside the TBD Lab unit of its Superintelligence Labs and unveiled in April to compete head-on with OpenAI, Anthropic and Google. The first slip, from April to May, was blamed on bugs found in testing and the need to build out more infrastructure. The target then moved to June. A company spokesman now says the interface is in partner testing and will ship this month, though he would not name a day.

Why the API matters more than the model

For a closed model, the developer API is not a side feature. It is the whole business. Outside companies cannot build Muse Spark into their own products without it, which means every week of delay is a week Meta earns nothing from the model and cedes ground to rivals whose tools are already live. As one summary of the situation put it, Meta's problem is no longer just model quality. It is execution, infrastructure and the trust of the developers it is asking to commit.

The delay lands against an eye-watering spending backdrop. Meta has raised its 2026 capital-expenditure forecast to as much as $145 billion, up from a January range of $115 billion to $135 billion, with much of that going to AI data centres. The company is pouring money into the buildout at the same time that its headline model cannot reach the developers meant to use it.

Meta is not the only large company trying to reduce its dependence on outside labs. Microsoft recently rolled out seven of its own in-house models to lean less on OpenAI. The difference is that those shipped. Muse Spark, for now, is a launch announcement waiting for a launch.

None of this means the model is bad. Meta has not cited any performance or safety problem, only the unglamorous work of fixing bugs and standing up infrastructure. But in a year when a rival ships a new frontier model seemingly every few weeks, a flagship that keeps missing its own dates starts to tell a story of its own.