Meta introduces Muse Spark 1.3 for longer agentic and coding tasks
Meta AI Research introduced Muse Spark 1.3 on September 2, 2026, as an update aimed at agentic workflows and coding. The model with maximum reasoning is available in Muse Code and Meta Model API. Unlike a simple chat-model refresh, Meta frames the release around longer tasks: an agent should gather context from messy sources, manage several workflows in one thread, correct gaps in its plan and deliver a finished result. Those capabilities are described by Meta, not independently verified performance guarantees.
The release also changes how the agent is expected to work with a person. Meta says Muse Spark 1.3 can ask clarifying questions when an instruction is ambiguous, ask for help when it is stuck and confirm before taking consequential actions. It is designed to adapt its level of updates to the user, either reporting frequently or continuing quietly. Meta further says it follows complex, long-form instructions more reliably and is better at mapping an incoming request to the right task when a single conversation contains interruptions and older work.
Coding is the clearest quantified claim in the announcement. In comparisons by Meta engineers, Muse Spark 1.3 used about 20 percent fewer tool calls and 25 percent fewer tokens than Muse Spark 1.2. The comparison is useful as a signal about efficiency, but it is not an independent benchmark and does not show how the model performs on every codebase, toolchain or budget.
Meta also reports improvements in adversarial robustness, prompt-injection resistance and calibration around irreversible actions. The company says the model has a better sense of its limits and is less likely to claim success when it encounters a hurdle. Those are important design goals for agents with access to files, software and external tools, but real-world safety still depends on permissions, logging, testing and human approval.
For developers, the practical news is immediate hosted access rather than a future preview. Teams can test Muse Spark 1.3 through Meta’s two developer surfaces and compare quality, latency, cost and failure recovery with existing models. Meta says an open-weights release and larger models are on its roadmap; neither is part of today’s availability. The update therefore strengthens Meta’s push toward long-running software agents while leaving users to verify whether the promised efficiency and discretion survive real workflows.