Three months after our July 11 coverage of Muse Spark 1.1, Meta has released two Muse Wonder models aimed at doing agent-style work.
Meta calls the pair its most capable models yet for tasks where the system must read, reason and act: helping people navigate insurance claims, analyzing leases, and even building applications. The family includes Muse Wonder, available through the Meta Model API and Muse Code, and Muse Wonder Full, which Meta is making available to Muse Code subscribers. Meta says Muse Wonder will soon power features in its Muse assistant, with subscribers getting first access.
Why this matters: Meta is pushing beyond research demos toward models it believes can actually run real multi-step assistant jobs. That capability claim is the main signal here, and Bloomberg reports Meta describes Muse Wonder as competitive with leading models from other labs, a company claim, not independent proof.
How it works, at a high level: these models are built to handle multiple kinds of inputs and follow through on tasks, not just answer a single question. Think of them as an expert who can read a contract, pull the important points, and then draft the right response. That design makes them useful for multi-step, professional workflows.
What changes now is practical. Developers get two access paths: the hosted API for wider integration and Muse Code for deeper builds, while subscribers can try the fuller model first. That lowers the barrier to building agentic features, but it does not make frontier models trivially cheap to run, teams will still rely on hosted services or serious infrastructure.
What's next: Meta's claim of frontier-level capability will stand or fall on independent benchmarks and real-world apps. The open question is whether outside teams will turn this access into products that genuinely compete with the closed systems already dominating the market.
