Meta releases Muse Spark 1.3 with stronger agentic workflows and leaner coding

What shipped

Meta has released Muse Spark 1.3, an update to its model lineup that the company says performs better on agentic and coding work while being easier to run in real-world conditions. According to Meta's announcement on its research blog, which reached the front page of Hacker News, the release draws on months of feedback from broad use of Muse Code and the Meta Model API.

The new model is available now in Muse Code and through the Meta Model API. Reasoning modes carried over from earlier releases are live from day one, while a maximum reasoning mode will arrive later, after Meta finishes additional safety testing.

Agentic behaviour

Meta's central claims concern long-running agent work. The company says that when handed an open-ended objective, Muse Spark 1.3 uses tools to assemble its own context from messy or contradictory material, spots and repairs gaps in its own plan, and keeps a running record of what it has learned until it produces a finished deliverable. To keep that behaviour from being tied to a single environment, Meta says it trained the model across a diverse set of agent harnesses.

The model is also tuned to behave more like a collaborator. Meta says it asks clarifying questions when a request is vague, calls on the user when it gets stuck, and seeks confirmation before taking consequential actions. On lengthy jobs it adjusts to the user's preferred style, either reporting progress frequently or working quietly in the background.

Instruction following receives specific attention. Compared with earlier Muse Spark models, version 1.3 is said to track complex, long-form instructions more reliably across multi-step tasks, holding onto detailed requirements without dropping constraints or drifting away from the requested workflow. Multitasking improves as well: the model more accurately routes each incoming prompt to the right task inside a cluttered single-threaded conversation, even when the user interrupts or redirects earlier requests.

Meta also claims better self-knowledge. The model is trained to recognise what it can and cannot do, what it actually knows, and when it has hit an obstacle, rather than inventing outcomes.

Coding improvements

For developers, the standout numbers are on efficiency. Meta says 1.3 was trained on more long-horizon coding tasks and behaves better in common engineering workflows. In comparisons run by Meta's own engineers, it came out significantly faster and more efficient than Muse Spark 1.2, using roughly 20% fewer tool calls and roughly 25% fewer tokens. Meta also describes fewer unnecessary turns, less verbosity and a cleaner overall coding style.

To illustrate the agentic side in practice, Meta's post includes an example in which the model drafts a flow-simulation report from preliminary CFD results and a CAD file, organised into standard report sections and exported as a PDF.

Safety and roadmap

On safety, Meta points to stronger resistance to adversarial inputs and prompt injection, along with better calibration on complex agent tasks about which actions are irreversible. The company frames both changes as improved judgment in long-running work.

Looking ahead, Meta says its roadmap includes larger models and an open-weights release for Muse Spark, in addition to the still-pending maximum reasoning mode.

Why it matters

Agentic capability is where frontier model competition is currently concentrated, and Meta is staking its claim on practical efficiency rather than raw benchmark scores. If the roughly 20% reduction in tool calls and 25% reduction in tokens holds up outside Meta's internal tests, it translates directly into lower cost and latency for agent-driven development, which is what many buyers actually feel. The caveats are real: the efficiency comparisons come from Meta's own engineers rather than independent evaluation, and the strongest reasoning mode is gated behind unfinished safety work. An open-weights release, if it materialises, would place this model line in direct competition with other open-weight offerings and give developers a freely deployable alternative to closed agent platforms.