Meta Stays Open: Strong Models Released, Frontier Models Stay Closed
Meta announced the release of Muse Spark Open, claiming it to be the most capable open-weight model in the US, while confirming that its more powerful frontier model, Muse Wonder, will remain closed. Meta's Chief AI Officer Alexandr Wang stated that the company will not open-source its absolute top-tier models due to safety and capability boundary considerations. This policy clarification highlights a strategic shift among major AI developers, balancing the benefits of open-weight models with the commercial and safety protections of proprietary frontier systems. It signals a more cautious approach to open-source AI deployment compared to previous absolute advocacy. Muse Spark Open performs below the recently announced Muse Wonder model, which reaches frontier levels without public benchmark scores. This follows Meta's previous open-source releases, including the Apache 2.0-licensed Glimmer model distilled from Muse Spark.
## BACKGROUND
Open-weight models refer to artificial intelligence models whose learned parameters, such as weights and biases, are publicly released for users to download and use. Model distillation is a technique where a smaller model is trained using the knowledge learned by a larger model, allowing the smaller model to inherit strong generalization capabilities.