Meta AI Released Muse Spark 1.3: An Agentic Coding Model That Uses ~20% Fewer Tool Calls and ~25% Fewer Tokens Than Muse Spark 1.2

This week, Meta Superintelligence Labs launched Muse Spark 1.3. It is the fourth Muse Spark launch in 5 months, and the goal is long-horizon agentic and coding work somewhat than single-turn technology. The framing in Meta’s submit is usability: sustaining an extended thread, collaborating with the person, and understanding when it’s caught.

Is it deployable? Yes, however with two limits. Muse Spark 1.3 ships at this time in Muse Code and the Meta Model API, so you may name it in manufacturing now. You can not self-host it, as a result of the weights are closed, and the max reasoning mode remains to be gated behind additional security testing.

What truly modified for brokers

Meta skilled Muse Spark 1.3 throughout a number of agent harnesses so conduct generalizes previous one surroundings. The mannequin is constructed to carry a number of workflows inside a single lengthy thread. Given an open-ended goal, it gathers its personal context from messy and conflicting sources, then patches gaps in its plan.

The collaboration modifications are the extra sensible half. Muse Spark 1.3 asks clarifying questions on ambiguous prompts, pulls the person in when it stalls, and confirms earlier than consequential actions. On lengthy runs it adapts to choice: frequent standing updates, or silent background execution. Meta additionally reviews higher calibration on the mannequin’s personal limits, so it flags hurdles as an alternative of hallucinating an end result.

Multitasking improved too. Meta says the mannequin maps an incoming immediate to the right process inside a cluttered single thread, whether or not the person is steering or interrupting.

Coding and effectivity

Muse Spark 1.3 was skilled on extra long-horizon coding duties. Relative to Muse Spark 1.2, Meta describes fewer pointless turns, much less verbosity, and a cleaner code fashion. In inside comparisons by Meta engineers, it used roughly 20% fewer device calls and roughly 25% fewer tokens. For agentic workloads, that’s the quantity that maps to price: fewer spherical journeys and fewer billed tokens per accomplished process.