Muse Spark 1.3 handles longer tasks with fewer tool calls

Meta releases Muse Spark 1.3 for coding and long-running agent workflows, but its strongest reasoning mode and promised open weights are not yet broadly available.

An AI agent receives simulation data and a CAD file, examines the material, writes an aerospace engineering report, organizes the measurements into tables, and exports the result as a PDF. Meta uses this example to illustrate Muse Spark 1.3, a model designed to keep working through complex assignments without losing the user’s original requirements.

Released on September 2, Muse Spark 1.3 succeeds the 1.2 version after only a few weeks. It is available through Muse Code, Meta’s coding environment, and the Meta Model API. The update focuses as much on the way the model conducts a task as on the quality of its final response.

Meta says Muse Spark 1.3 can manage several workflows within a single, lengthy conversation. It is intended to associate each new instruction with the correct task even when the user interrupts an ongoing request, returns to an earlier one, or changes the priorities of several assignments.

For an open-ended objective, the model can use tools to gather its own context from documents and sources that may be incomplete or contradictory. It is also trained to identify gaps in its plan, retain what it has already learned, and revise its approach before producing the requested deliverable.

The company has paid particular attention to interactions between the agent and its user. When a request is ambiguous, Muse Spark 1.3 is supposed to ask for clarification. When it encounters a problem it cannot resolve, it can request help instead of claiming that the task was completed. It should also obtain confirmation before taking an action that could have significant or irreversible consequences.

These behaviors do not mean that the model can reliably determine the consequences of every action. Whether it requests confirmation still depends on its interpretation of the situation, the instructions supplied by the application, and the permissions granted to its tools. An application integrating the model must therefore continue to control access to external systems and independently approve sensitive operations.

Progress reporting can be adjusted to the user’s preference. Muse Spark 1.3 may provide regular updates during a long assignment or operate more quietly until it reaches a result. This flexibility is particularly relevant to tasks that involve numerous searches, file transformations, terminal commands, or revisions.

Meta also says the model follows lengthy instructions more consistently than earlier Muse Spark releases. It should be less likely to abandon a formatting rule, overlook an expected file, or drift away from the prescribed process after several intermediate steps.

The demonstrations cover a wide range of professional work. In one example, the model receives preliminary computational fluid dynamics results and a STEP file representing an experimental wing. It prepares a report containing the simulation environment, boundary conditions, performance measurements, an aerodynamic discussion, and design recommendations.

Other examples involve correcting errors in a recorded bass track, preparing a presentation for a county advisory board, and summarizing public feedback from a spreadsheet. Meta explicitly describes these demonstrations as agent prototypes rather than finished products. They show possible workflows, but the company does not publish detailed success rates for each example or indicate how much human review was needed before presentation.

Coding remains one of the central uses of Muse Spark. Version 1.3 was trained on more long-running software engineering assignments and is designed to avoid unnecessary exchanges. Meta engineers found that it used approximately 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 during their comparisons.

Those figures could translate into shorter execution times and lower costs when an agent repeatedly searches a repository, edits files, runs tests, and corrects failures. They are nevertheless internal measurements. Meta does not disclose the number of projects involved, their difficulty, the exact reasoning configurations used, or a reproducible test allowing outside teams to confirm the savings.

Independent measurements also show that efficiency depends on the workload. Artificial Analysis estimated that the publicly available xhigh configuration cost about $0.55 per task on its Intelligence Index, compared with $0.40 for Muse Spark 1.2. This does not directly contradict Meta’s coding comparison, because the independent evaluation covers a broader mix of reasoning and agentic work. It does show that fewer calls in one engineering workflow do not guarantee lower consumption in every application.

The principal qualification concerns the benchmark configuration. The scorecard displayed in Meta’s announcement mainly reports Muse Spark 1.3 with max reasoning. That setting was not part of the broad release at launch. Previously available modes, including xhigh, can be used immediately, while max is expected after additional safety testing.

The difference is material on some evaluations. Meta reports an Elo rating of 1,754 for max on GDPVal-AA v2, which measures the production of professional deliverables. The available xhigh configuration reaches 1,709. On JobBench, the respective scores are 64.9 and 61.2.

The gap is larger on OSWorld 2.0, an evaluation in which an agent operates a graphical computer environment. Max scores 66.9, compared with 57.2 for xhigh. An organization assessing the currently deployable service should therefore not assume that it will obtain the score highlighted for max.

Other results show little difference between the two configurations. Both reach 89.4 on DeepSearchQA. On Terminal-Bench 2.1, which tests work performed in isolated terminal environments, xhigh scores 89.2 while max reaches 88.8. Additional reasoning effort does not automatically improve every task.

The max version performs strongly on Meta’s main coding comparisons. It records 75.4% on DeepSWE v1.1, ahead of the 74% attributed to Opus 5 and the