Meta Muse Spark
Meta positioned Muse Spark as the first model in the new Muse family from Meta Superintelligence Labs (MSL), led by chief AI officer Alexandr Wang, and framed it as a step toward 'personal superintelligence.' The company described it as a natively multimodal reasoning model with tool-use, visual chain-of-thought, and multi-agent orchestration, including a 'Contemplating' mode that runs sub-agents in parallel to rival extreme-reasoning systems. Meta highlighted leading scores on health benchmarks and competitive results on visual STEM and agentic tasks. Most notably, Meta presented Muse Spark as proprietary and closed-weight, a deliberate departure from the open Llama lineage.
Muse Spark shipped as a closed-weight, hosted model with no public weights, accessible to consumers through the Meta AI app and meta.ai, plus a private API preview for select users rather than a general developer API. Meta offered it free at the consumer tier and did not publish API pricing or a broad availability timeline at launch.
Two-plus months on, the durable story is strategic, not technical: Meta closed the weights on its frontier model, ending the open-by-default Llama era and conceding that catching OpenAI, Google, and Anthropic mattered more than its open-source posture. On independent indices Muse Spark landed credibly mid-pack rather than at the very top, trailing the leading GPT, Gemini, and Claude tiers while genuinely leading on health benchmarks. The 'Contemplating' parallel-agent mode is a real capability but echoes patterns competitors already shipped, so it reads as catch-up rather than a leap. For a company whose AI brand was built on openness, the quiet part is loudest: the moat is now the product and distribution, not the model. Executives should read this as Meta repositioning to compete on consumer reach and verticals like health, not on open-weight goodwill.
Meta abandoning open weights for its flagship signals that even the loudest open-source champion now treats frontier models as proprietary assets, narrowing the field of credible open alternatives enterprises can self-host.