Meta pivots from open Llama toward proprietary models with Muse Spark

Meta has launched Muse Spark, its first new model since the formation of Superintelligence Labs — and, notably, a proprietary one. It's a marked departure from the open-weight approach that made the Llama series a default choice for teams that wanted a capable model they could self-host and fine-tune without a licensing negotiation.

The shift has been building for a while. Llama 4 Scout and Maverick, a 400-billion-parameter open-weight model, shipped to broad availability on Hugging Face, and a larger teacher model, Llama 4 Behemoth, was trained to outperform GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-heavy benchmarks — but it's still training and hasn't been released. Meanwhile, Meta's next-generation model, internally called Avocado, has been delayed into 2026 alongside a broader reorganization toward closed-source development.

Why it matters for teams building on Llama

Meta says it still plans to release open-weight versions of some future models, so this isn't necessarily the end of open Llama releases outright. But teams that have built infrastructure around always-open Llama weights should treat that as no longer guaranteed by default, and plan model-sourcing decisions accordingly rather than assuming the next flagship ships open.

Source: Digitimes. VentureBeat