Meta's Muse Wonder goes frontier and closed, ending the Llama open-weights era

Meta's frontier AI model business: training proprietary models and selling access through an API and its own products.

That Meta's frontier models are worth more closed — sold through an API and its own paid coding tool — than given away as downloadable open weights like Llama.

Meta built its AI reputation on Llama, the downloadable open-weights family that organisations ran on their own infrastructure. In or around August 2025 Zuckerberg reorganised the effort into Meta Superintelligence Labs under Alexandr Wang, recruiting research talent and committing the Prometheus AI datacenter.

On October 12, 2026 Meta announced Muse Wonder and Muse Wonder Full: trained on Prometheus with large-scale pretraining plus reinforcement learning, multimodal input, roughly one-million-token contexts — and no downloadable weights for either. The standard Wonder sells through a commercial API at $4/$20 per million input/output tokens (cached input $0.20); Wonder Full is reserved for paid Muse Code subscribers. Artificial Analysis scored Wonder Full 56 on its Intelligence Index — level with Claude Sonnet 5.5, ahead of GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 (53), behind only Claude Opus 5.5 (58); standard Wonder scored 52. Wonder Full leads SWE-Marathon long-running coding at 57.5 vs 46.3 for Opus 5.5, but trails Opus 5.5 and GPT-6 Astra on Terminal-Bench 4.0 and Opus on HealthBench Pro.

The Llama strategy commoditised models and won Meta goodwill but captured little frontier revenue; the API at $4/$20 per million tokens monetises the lead directly.

Reserving Wonder Full for paid Muse Code subscribers ties the best capability to Meta's own product surface instead of powering rivals' applications.

Independent benchmark placement was the launch's proof point: Artificial Analysis's scores, cited in Meta's own comparison, put Wonder Full at the frontier on the day it shipped.

The benchmark mix favours the pitch — SWE-Marathon's long-running coding tasks are exactly where Active Memory and Active Focus should show, and Wonder Full leads by 11 points.

Strategic assets get walled when they lead: Meta open-sourced Llama while behind, then closed Wonder once third-party scores put it at the frontier.

As of the October 12, 2026 announcement, standard Wonder is available through the API and Wonder Full only via paid Muse Code; neither has downloadable weights, closing the door on the distribution model that made Llama ubiquitous. VentureBeat flags the open questions for enterprise buyers — cost, data governance, and whether Wonder Full will ever be available for production deployments outside Meta's tools — while Claude Opus 5.5 holds the top aggregate score, keeping the frontier contest open.