Meta’s new Muse Spark model marks the company’s first major break from its open-source AI strategy built around Llama.
Meta has launched Muse Spark, the first model from Meta Superintelligence Labs. It is also the company’s first proprietary large language model. The launch ends a three-year open-source strategy built around the Llama family.
Muse Spark is proprietary. Meta says it hopes to open-source future versions, but offers no timeline. The model is the first product from the unit that Alexandr Wang leads. Zuckerberg recruited Wang after growing frustrated with Llama’s progress against OpenAI’s ChatGPT and Anthropic’s Claude. To bring Wang in, Meta invested $14.3 billion in Scale AI for a 49% stake.
Why It Matters
Muse Spark supports tool-use, visual chain of thought, and multi-agent orchestration. It is natively multimodal from the ground up. According to Meta, improved training techniques and rebuilt infrastructure now allow the company to build smaller models that match Llama 4 Maverick’s capability for an order of magnitude less compute. The model already powers the Meta AI app. Rollout to Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban AI glasses follows in the coming weeks.
However, the business shift matters more than the benchmarks. Meta now plans to offer third-party developers API access to Muse Spark – a direct revenue stream the company never pursued with Llama. That move signals a fundamental change in strategy. For the AI Business sector, it suggests open-weight distribution is no longer viable once a company believes it has a model worth protecting.
Meanwhile, the competitive pressure from China helps explain the timing. By late 2025, Chinese models from Alibaba and DeepSeek accounted for 41% of downloads on Hugging Face. New entrants like Zhipu AI’s GLM-5 and Alibaba’s Qwen 3.6 Plus had already outpaced Llama 4 Maverick on key benchmarks. As a result, the open-source ecosystem Meta built is now being outcompeted – removing much of the strategic case for continued openness.
What’s Next
Meta plans to spend between $115 billion and $135 billion on AI infrastructure in 2026. That is nearly twice last year’s capex. Still, that level of spending requires a revenue model that free model distribution simply cannot support. The API preview and planned commerce integrations are the first concrete steps toward closing that gap.
Yet the open-source developer community faces real uncertainty. Wang’s comment that future versions “may be open-sourced” carries no commitment and no date. Furthermore, Muse Spark launches with tighter access restrictions than even the paid proprietary models from OpenAI and Anthropic. That is a remarkable position for a company that built its AI credibility on openness.
In the end, Muse Spark’s success will not be decided on benchmark tables. It will be decided across Meta’s 3+ billion user surfaces – on WhatsApp, Instagram, and Facebook. The deployment advantage is unmatched. The question is whether everyday users notice the difference.
Sources: Meta AI · CNBC · VentureBeat · Fortune
About the Author
