Meta released AI models Muse Wonder and Wonder Full

The company Meta presented two new models of artificial intelligence - Muse Wonder and Muse Wonder Full. This was reported by the company's representatives.

Both closed models are focused on long-term agent scenarios, within which the system autonomously solves multi-step user tasks, including programming, working with documents and personal finances. The new models have implemented Active Memory functions to maintain context between dialogues and Active Focus to adapt to errors and continue working on the task. The training of the systems was carried out in the Prometheus data center with a capacity of about one gigawatt based on the Muse Spark architecture.

According to Meta's internal benchmarks, in the DeepSWE 1.1 software development task test, the Wonder Full model scored 79 points, outpacing Claude Opus 5.5 (74.2) and GPT-6 Astra (74.1). In the SWE-Marathon test, the older model scored 57.5 points against 46.3 for Opus 5.5 and 42.5 for Astra. At the same time, in the independent Artificial Analysis Intelligence Index rating, Muse Wonder Full lost to Claude Opus 5.5, scoring 56 points against 58, but outpaced GPT-6 Astra (53 points).

Access to the basic Muse Wonder is open through the Meta Model API and the Muse Code service for developers. The older Muse Wonder Full is provided to Muse Code subscribers. In the future, Meta plans to integrate the new model into the Muse assistant, opening access first to paid subscribers.