Zuckerberg ends a 3-year X silence to launch Meta's Muse Spark 1.1

Mark Zuckerberg posted on X for the first time in three years, and it was an AI model launch.

His previous post dates back to July 2023. On July 9, 2026, he returned to Elon Musk’s platform to announce Muse Spark 1.1, Meta’s low-cost agentic and coding model.

What Muse Spark 1.1 is

The model targets the practical end of the agent market. The spec sheet, per Meta:

• A 1 million token context window
• Parallel sub-agents for splitting work
• Training to operate desktop, mobile and browser interfaces

That last item matters: a model trained to click through real interfaces is aimed at doing tasks, not just discussing them.

On benchmarks, Meta claims 88.1 on MCP Atlas, which measures scaled tool use, and 54.7 on JobBench. The company says Muse Spark 1.1 rivals much pricier frontier models on agentic evals. Those are Meta’s own numbers, so independent verification will tell the real story, but the positioning is unambiguous: frontier-adjacent agent performance at a budget price.

More than a model launch

Two additional pieces shipped alongside the announcement. A public preview of the Meta Model API launched the same day, giving developers direct access to Meta’s models as a service. And in September, Meta starts manufacturing its own AI chip, Iris.

Taken together, that is a vertical stack taking shape: Meta’s own silicon, its own API, and a cheap agentic model to run on top.

Cheap agents are the next battleground

The strategic read is straightforward. The frontier model race is expensive and crowded. The market for agents that are cheap enough to run constantly, on every task, at scale, is just forming, and Muse Spark 1.1 is Meta’s play for it.

As for announcing it on X: choosing your rival’s platform for your first post in three years, to launch a product that competes with your rival’s AI company, reads however you want it to read. Petty or genius, it got everyone’s attention.