Meta Releases Muse Spark 1.3 for Coding Agents
Muse Spark 1.3 targets sustained coding and agent workflows, while Meta postpones both open weights and its highest reasoning setting.
An agent-focused model update
Meta released Muse Spark 1.3 through Muse Code and the Meta Model API, positioning the update around sustained coding, tool use and real-world agent tasks. The release is immediately usable as a hosted model, but it is not yet an open-weight release: Meta said downloadable weights will follow later, without giving a firm publication time.
The rollout also separates the model’s currently available reasoning settings from its most compute-intensive configuration. Meta said the “max” reasoning mode will arrive only after additional safety testing. That distinction matters because some of the strongest third-party results disclosed on launch day came from the unreleased max setting, rather than the configuration developers can deploy immediately.
Artificial Analysis reported that Muse Spark 1.3 max scored 52% on Tau3-Bench Banking, the highest result it had measured on that evaluation. The available xhigh setting reached 47%, while the evaluator described the model as unusually cost-efficient at its intelligence level. Meta kept API pricing at $1.25 per million input tokens and $4.25 per million output tokens. The reported improvements were concentrated in agentic evaluations rather than spread evenly across every category.
Availability tempers the headline numbers
Muse Spark 1.3 strengthens Meta’s attempt to compete in coding agents through both a first-party harness and an API, while preserving the prospect of a later local ecosystem around open weights. That staged strategy can generate adoption quickly without releasing the most sensitive or expensive configuration on day one.
Its significance will depend on what Meta ultimately publishes. Hosted access gives developers a practical new option now, but the open-source community cannot yet inspect, fine-tune or independently serve the weights. Likewise, results from the pending max mode should not be treated as the performance of today’s generally accessible product. The consequential milestone will be whether Meta delivers the promised weights with capabilities close to the hosted release and terms broad enough for serious deployment.