Meta ships Muse Spark 1.3 with max reasoning, closed weights

Meta released Muse Spark 1.3 and says max reasoning is now live on Muse Code and the Meta Model API (dev.meta.ai). The ship is a closed-weights coding and agent model you touch through Meta’s harness or API — not a Llama-style weights drop.

What it does: Meta says 1.3 holds longer-horizon agent work in a single messy thread: tool use across conflicting sources, plan repair, multitasking when the user interrupts, clarifying questions on ambiguous prompts, and confirmation before consequential actions. On coding, Meta engineers report ~20% fewer tool calls and ~25% fewer tokens versus Muse Spark 1.2, with less needless verbosity. Safety notes in the same post: stronger resistance to adversarial inputs and prompt injections, and better calibration on irreversible actions in agentic tasks.

What it cannot do (yet): You do not get open weights with this release. Meta’s “Looking Forward” line puts a Muse Spark open weights release on the roadmap with bigger models — undated, unspecified which model. Until that lands, the product is API/harness access only.

Who it is for: Developers already on Muse Code or Meta Model API who need a longer-horizon coding agent with max reasoning turned on. Teams that require self-hosted or inspectable weights are still waiting on the open release Meta has not scheduled.

What it replaces: Muse Spark 1.2-class workflows inside Meta’s own surfaces. It does not replace open Llama stacks for shops that refuse closed models.

Risk: A closed agent that asks clarifying questions and confirms before consequential actions is still an agent that can act with tools. Treat irreversible-action calibration as Meta’s claim, not an independent audit. Marketing toward “personal superintelligence” is not a measured capability.

What’s next: Dated open-weights announcement; whether max reasoning stays generally available or gets re-gated after more safety testing; Meta’s separate evaluation report numbers under outside scoring.

— Rowan Keel