Meta released Muse Spark, its first major AI model in a year and the first product from its newly formed Meta Superintelligence Labs—but this time, there's no open-source release. The natively multimodal model with tool use and multi-agent orchestration now powers Meta AI across three billion users, scoring fourth overall on AI benchmarks while leading in health applications with a 42.8 score on HealthBench Hard. Meta spent $14.3 billion acquiring Scale AI talent, brought in 28-year-old Alexandr Wang to lead the rebuild, and tore down its entire AI stack over nine months to create a model that runs at a fraction of Llama 4's compute cost.
This marks a fundamental shift for the company that built its AI credibility on open weights. Llama reached 1.2 billion downloads by early 2026, becoming the reference point for open-source AI development, but Meta is now adopting a "hybrid license" approach where frontier models stay proprietary. The pivot follows Llama 4's underperformance and concerns about competitors like DeepSeek exploiting Meta's open-source work to build competitive models—a direct challenge to Meta's previous strategy of using open source as a competitive moat against closed providers.
The developer community that made Llama successful is now left waiting for promised smaller derivatives that "may eventually get open weights," with no timeline commitments. Meta's reasoning centers on safety reviews and protecting model specifications, but the practical reality is that the company has joined OpenAI and Anthropic in keeping its most capable work behind closed doors. For developers who relied on Meta as the open-source alternative to proprietary giants, this represents a significant consolidation in the AI landscape—one fewer major player committed to open development at the frontier.