Meta releases Muse Wonder AI models built to run personal and professional agents
SiliconANGLE Duncan Riley ● Covered by 5 sources
Meta launched Muse Wonder, a pair of AI models for agents that handle long tasks. It says they’re its strongest yet, with memory and recovery features built in.
Based on reporting by SiliconANGLE, Duncan Riley — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta has pushed out Muse Wonder and Muse Wonder Full, a new pair of AI models aimed at agents that don’t just answer a prompt and stop. The company wants them on the kind of chores people hand off and forget about until something goes sideways: insurance claims, coding work, legal tasks, finance, health. Long jobs, in other words. The models are also meant to keep going when the first plan fails, which is the part that matters more than the shiny demo.
Two features drive that pitch. Active Memory carries context from one conversation to the next and brings details back when they become relevant. Active Focus is supposed to help the model push through hard stretches by trying another route, recovering from errors and checking with the user before doing anything important. Meta says the models can take in text, video, audio and images across a 1 million-token window, but they return text only.
The company trained both models at its gigawatt-scale Prometheus data center, building on the pretraining and reinforcement learning approach it used with Muse Spark in April. In the model card, Meta lined Muse Wonder Full up against OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Claude Opus 5.5. The comparisons were not all equally neat, though. Meta says its own runs of the competing models were best-effort, and the prompts and tools may not have been tuned for them.
Still, some of the numbers are hard to ignore. On Artificial Analysis’ version of Harvey’s Legal Agent Benchmark, Muse Wonder Full passed 13.1% of tasks, ahead of GPT-6 Astra’s 8.6%. In coding, the picture splits. Muse Wonder Full led DeepSWE 1.1 and SWE-Marathon, where it scored 57.5% against Claude Opus 5.5’s 46.3%, though Meta says that SWE-Marathon comparison was not controlled. On ProgramBench, which tests whether an agent can rebuild a full program from scratch, Opus 5.5 won 70.5% to 59%.
Meta is also trying to make the safety case before the full report lands. The model card says Muse Wonder Full could hit the high-risk threshold for chemical and biological threats if used without mitigations, but with those safeguards in place it rates moderate or lower across chemical and biological threats, cybersecurity and loss of control. On one agentic misalignment test, its harmful-action rate was 0%. Developers can use Muse Wonder now through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Meta says its Muse agent will get the model soon, with subscribers first.
My take — AI-written commentary, not fact-checked reporting
Meta loves to market restraint as intelligence, which is adorable when the product still has to prove it won’t wander into a locked room and start improvising. The real contest here isn’t who wins a benchmark by a few points; it’s who can build agents that stay useful without turning every workflow into a liability review. Open models keep winning trust because closed ones keep asking for it like a favor.
Read more about this at: SiliconANGLE
