Meta Releases Open Weights for Muse Spark While Keeping Frontier Model Closed

Meta has published the weights for Muse Spark Open, a model designed for coding and autonomous agents that the company acknowledged is deliberately less capable than its flagship frontier system, according to a report by The Next Web (https://thenextweb.com/news/meta-muse-spark-open-weights-eu).

The release formalizes a strategic shift for Meta. Chief AI Officer Alexandr Wang stated that the company will generally not open-source its most powerful models due to safety and capability considerations, while maintaining that Meta intends to continue releasing capable open models below the frontier threshold.

Muse Spark Open marks the second open release in the Muse series. In August, Meta released Glimmer, a 30-billion-parameter model distilled from Muse Spark under an Apache 2.0 license, developed with Apple to run on dual M5 Ultra Mac Studios. After deliberating in September over whether to publish weights for Muse Spark 1.3, Meta opted to release Spark Open across Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Nebius, the Dutch infrastructure provider to which Meta has committed up to $27 billion.

By contrast, Meta's proprietary Muse Wonder model, announced on Monday without published benchmarks, will remain closed. This tiered distribution structure maps directly onto Article 53 of the European Union's AI Act, which exempts genuinely free and open models from AI Office documentation requirements unless they cross the threshold of systemic risk.

The policy change contrasts with remarks made five weeks earlier by Meta Chief Executive Mark Zuckerberg on the Sources podcast, where he argued that competitors training advanced models without releasing them publicly was dangerous. Meta has also declined to sign the EU code of practice operationalizing AI Act compliance, following executive Joel Kaplan's statement last year that European regulation was on the wrong path. Non-compliance with the AI Act carries potential fines of up to 3% of global turnover along with pre-release evaluation mandates.

The open-source AI landscape remains caught in broader geopolitical competition. Wang asserted that the majority of global AI workloads will run on open models and argued those models should be American, while expecting Chinese developers to build on Muse Spark Open. Wang's comments follow an address in Turin eight days earlier by Alibaba Chairman Joe Tsai, who told attendees that open source is Europe's only viable path to technological independence. Meanwhile, European competitor Mistral has introduced Large 4, a trillion-parameter open-weight model targeting China's lead in the category.