Meta Breaks From Open Source: Muse Spark Is Its First Frontier Model — and First Without Open Weights

Meta released Muse Spark on April 8 through its Superintelligence Labs division — the research unit the company formed after recruiting away several top AI researchers from Google DeepMind and other labs. The model is multimodal, supports tool use, features visual chain-of-thought reasoning, and includes "Contemplating Mode," a parallel agent processing architecture that lets the model decompose complex tasks across simultaneous reasoning chains. It scores 52 on the Artificial Analysis Intelligence Index, putting it in the top five overall and closing the 18-point gap that separated Meta's previous best model, Llama 4 Maverick, from the current frontier leaders.

The Open Weights Reversal

The more significant news may be what Muse Spark is not: open source. Meta has spent the past three years making open-weight releases of the Llama family its primary competitive differentiator against OpenAI and Anthropic. The argument was strategic as well as ideological — open models build ecosystem, ecosystem builds developer loyalty, developer loyalty converts to cloud and enterprise business. Muse Spark abandons that strategy for its frontier tier. The model is available only through meta.ai and a private API preview. Meta says it may release open-weight versions of future Muse models, but that commitment is deliberately vague.

Benchmark Position

At 52 on the Artificial Analysis Intelligence Index, Muse Spark ranks behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6, but ahead of a substantial portion of the current frontier field. On Humanity's Last Exam — a benchmark designed to be resistant to training data contamination — it scores 50.2 in extended thinking mode, which is competitive with current frontier models. Token efficiency is a genuine strength: the model uses approximately 58 million tokens to complete the benchmark suite, comparable to Gemini 3.1 Pro's 57 million and substantially better than Claude Opus 4.6's 157 million. Acknowledged weaknesses include long-horizon agentic tasks and complex coding workflows, where it lags Claude Sonnet 4.6 and GPT-5.4.

What This Means for the AI Ecosystem

Meta's decision to go closed on its frontier model has immediate implications for the open-source AI ecosystem. The Llama model family has been the foundation for thousands of fine-tuned models, local inference projects, and enterprise deployments that needed capable open weights without OpenAI's pricing. If Meta's frontier capability is now closed-only, the open ecosystem will need to look elsewhere — to Mistral, to community fine-tunes of older Llama versions, or to the smaller open-weight models from Google and others. The message from Meta is that open weights are a market development tool, not a commitment. When frontier capability becomes genuinely valuable enough to monetize, the calculus changes.