Meta's Muse Wonder brings it back to the frontier

Wired reports that Meta's largest model now competes with the best available, and adds features aimed at keeping AI agents on track.

Meta's newest AI models, Muse Wonder and Muse Wonder Full, suggest that the company's expensive effort to rebuild its AI program is starting to pay off, Wired reports.

What Meta released

Meta says Muse Wonder Full has frontier capabilities, putting it among the most capable models available. The company described the pair as the next step on its scaling ladder and a major move toward what it calls personal superintelligence. Wired notes that the models include new features intended to keep AI agents on track during long tasks.

The models are available through an API and through Muse Code, Meta's AI coding tool. Meta says the strongest version will also power features across its own platforms, including its Muse personal agent.

How Meta got here

Meta was once a leader in open AI with its Llama models, the first frontier-class open-weight models from a major lab, which were widely adopted by researchers and startups. After several disappointing Llama releases, and as OpenAI, Anthropic and Google moved ahead, the company fell behind.

Last year, Mark Zuckerberg invested billions in a new Meta Superintelligence Labs and hired former Scale AI CEO Alexandr Wang to lead it. The effort has had turnover, with some researchers leaving quickly. The first model from the new lab, Muse Spark, arrived in April, and Meta has since also released open weights for Muse Spark Open.

Why it matters for builders

Another frontier-level provider means more choice and pricing pressure. Coding is clearly a priority, with Muse Code competing in the same space as tools like Claude Code and Cursor. If you rely on AI for development, it is worth testing Muse Wonder on your own repository tasks.

The focus on keeping agents on track is also notable. The main failure mode for coding and research agents is drifting off task during long runs. Features that help agents stay aligned with instructions over many steps can matter more in practice than small gains on single-question benchmarks.

As with any new model, compare it on your own workloads and keep your stack flexible so you can switch.

Meta's open model, Muse Spark Open, is listed in our New AI Models category.

Source: Wired. This article is PickAI’s own summary and analysis based on the original reporting.