Meta released Muse Spark 1.3 on Tuesday, a coding and agentic update to its Muse Spark model line, and shipped it into Muse Code and the Meta Model API immediately. One reasoning tier did not ship with it: max reasoning, the model’s highest setting, is held back while Meta finishes additional safety testing.
That split matters more than the release itself. Meta is treating “available” and “fully tested” as separate milestones, shipping the parts of a model it has cleared while keeping the most capable mode in evaluation. It is the same posture OpenAI and Anthropic have taken with their own top-tier reasoning configurations this year, and it is becoming the default release pattern at frontier labs: a model is no longer one artifact that ships or does not, it is a set of tiers that clear safety review on separate timelines. For a reader tracking capability, that means “Muse Spark 1.3 is out” and “Muse Spark 1.3’s ceiling is out” are now two different facts to check.
On the capability side, Meta says Muse Spark 1.3 is built to sustain longer stretches of agentic work inside a single thread. The model generates its own context from messy or conflicting sources, corrects gaps in its plan without a fresh prompt, and tracks what it has already learned toward a final deliverable. Meta also trained it to interrupt itself: asking clarifying questions when a request is ambiguous, flagging when it is stuck, and confirming before consequential actions rather than proceeding silently.
Meta reports the model follows complex, multi-step instructions more reliably than Muse Spark 1.2, preserving constraints across longer workflows instead of drifting from them partway through. Multitasking is also claimed to be better: when a thread gets messy and a user cuts across one request with another, the model is said to route each prompt to the job it belongs to.
On coding specifically, Meta says internal comparisons by its own engineers found Muse Spark 1.3 uses roughly 20% fewer tool calls and roughly 25% fewer tokens than Muse Spark 1.2 on the same tasks, while producing a cleaner coding style with fewer unnecessary turns. Those figures come from Meta’s own testing, not an independent benchmark, and the release does not include third-party verification of the efficiency claims.
Meta frames the safety work as an extension of the capability gains rather than a separate track. Meta also credits the release with holding up better against hostile inputs and injection attempts, and with weighing whether an action can be undone before an agent takes it. Meta has not published the methodology behind those safety claims beyond a linked evaluation report, so the specifics of how “better calibration on irreversible actions” was measured are not independently confirmable from the launch post.
Meta says a larger model and an open-weights release in the Muse Spark line are both coming, without a date. The max reasoning tier’s arrival is also undated beyond “shortly.”
Teams building agentic coding workflows on Muse Code should benchmark Muse Spark 1.3’s standard reasoning modes against their current stack now, but should not assume feature parity with whatever max reasoning eventually ships. That tier’s safety hold is Meta’s own signal that its behavior under harder tasks is not yet fully characterized.
Meta Superintelligence Labs detailed the release in a September 2, 2026 post on Meta’s research blog.
