Meta released Muse Spark 1.3 on September 2, 2026, saying the update improves agentic and coding work after months of use in Muse Code and the Meta Model API. The new model is rolling out in those two surfaces today. Previously available reasoning modes ship with it. A higher-effort “max reasoning” mode is waiting on more safety testing.

The lab frames 1.3 as a step toward personal superintelligence that is easier to use in real settings, not only a benchmark bump. It published a separate evaluation report alongside the post.

Longer threads, more interruptions

Meta says Muse Spark 1.3 is trained to keep a long, messy thread together. Given an open-ended goal, it is supposed to gather its own context from conflicting sources, fix gaps in the plan, and produce a deliverable. The company trained it across several agent harnesses so it would not overfit to one tool stack.

The model is also meant to collaborate more. Meta says it asks clarifying questions when a prompt is vague, asks for help when stuck, and confirms before consequential actions. On long jobs it can send frequent updates or work quietly, depending on what the user prefers. The lab says it follows long, detailed instructions more reliably and maps a new prompt to the right task when several jobs share one thread.

Meta also says 1.3 has a better sense of what it can and cannot do, and is less likely to invent a finished outcome when it hits a wall. Those claims sit in the blog. The numbers live in the evaluation report.

Decoded Take

Muse Spark 1.3 is Meta admitting the first Muse drop was a product surface looking for a workhorse. Shipping in Muse Code and the Model API, while holding max reasoning for safety, is the same pattern OpenAI and Google are using: show the agent, park the sharpest mode. The test is not whether 1.3 asks nicer clarifying questions. It is whether a team will keep a multi-hour coding or research job in one Meta thread instead of jumping back to Claude or Gemini. Watch when max reasoning actually appears, whether the evaluation report names public benchmarks or only internal harnesses, and whether Muse Code usage is still a Meta-internal loop or a thing outside companies will pay for.