Meta has released Muse Wonder and Muse Wonder Full, the models previously known by the codename Watermelon.[4] The standard model is available through Meta’s API and Muse Code, while the stronger Full version is currently restricted to Muse Code subscribers.[1]
Independent testing supports Meta’s claim that it has reached the frontier, although not that it leads every category.[2] Artificial Analysis gave Wonder Full 56 on its Intelligence Index, up eight points from Meta’s Muse Spark 1.3 and within the leading group of current models.[2] Wonder Full placed in the top three on six of the index’s ten evaluations, with particular strength in agentic knowledge work, scientific reasoning and coding agents.[2]
The gaps still matter.[2] Artificial Analysis found knowledge to be a relative weakness: Wonder Full scored below Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon on its knowledge-focused AA-Omniscience evaluation.[2] Meta’s own model card also warns that performance can vary with prompts, tools, serving configuration and the surrounding agent system.[1]
What makes the launch more significant is the machinery behind it. Meta says both Wonder models were trained at Prometheus, its gigawatt-scale data-centre system.[1] Meta’s engineering team describes Prometheus as a cluster spanning several data-centre buildings and interconnecting tens of thousands of GPUs through a high-capacity network.[3] Its backend aggregation layer is designed to connect regional fabrics with Meta’s wider backbone, with inter-region capacity measured in tens of petabits per second.[3]
That infrastructure sits inside a company that already operates global consumer platforms used by billions of people.[1] Meta therefore has more than a competitive model: it has the compute to keep training and serving successors, established products through which models can reach users, and an internal agent product in Muse.[1][4]
There are limits to the comparison. Wonder Full is subscription-gated rather than generally available through an API, making direct independent testing and deployment harder than for the standard model.[1][2] Axios also noted that Meta had not said whether it would release the models’ weights, marking a break from the open-weight strategy associated with Llama.[4]
Muse Wonder does not settle the frontier race.[2] It does show that Meta is no longer competing only through reach or open models: it now has an independently measured top-tier system and the industrial infrastructure to keep pressing the leaders.[2][3]
[1] Meta AI Research: Muse Wonder
[2] Artificial Analysis: Muse Wonder and Muse Wonder Full
[3] Engineering at Meta: Building Prometheus
[4] Axios: Meta unveils frontier-class AI models
Source: Meta AI Research