Meta on Wednesday launched Muse Spark, the first AI model produced by Meta Superintelligence Labs under Alexandr Wang, nine months after the company’s $14.3 billion deal to bring in the former Scale AI CEO. The model is now live in the Meta AI app and website in the US, with rollout to WhatsApp, Instagram, Facebook, Messenger, and Meta’s smart glasses planned for coming weeks, according to CNBC and The Verge.
Proprietary, Not Open-Source
The most notable shift: Muse Spark is proprietary. Meta built its previous AI reputation on open-source Llama models, which power a significant share of self-hosted agent deployments and local inference setups. Muse Spark breaks that pattern. The company said it has “hope to open-source future versions of the model,” per CNBC, but Muse Spark itself ships closed.
Select partners currently have access to a “private API preview.” Meta plans to offer paid API access to a broader developer audience later, creating a new revenue stream that Llama’s open-weight distribution never generated.
Performance Positioning
Meta is not positioning Muse Spark as a frontier model. The company describes it as “small and fast by design,” highlighting competitive performance in multimodal perception, reasoning, health, and agentic tasks rather than claiming top benchmark results. Meta’s technical blog states that improved training techniques allow the model to match its older midsize Llama 4 variant “for an order of magnitude less compute.”
The model supports multimodal input (text and images), offers both a fast “Instant” mode and a slower “Thinking” mode for deeper reasoning, and can run multiple AI sub-agents to handle complex queries, according to The Verge. Meta highlighted health as a focus area, showing the model estimating calorie counts for meals, a competitive move against OpenAI’s ChatGPT Health and Anthropic’s Claude for Healthcare.
The Competitive Gap
Meta’s AI capex for 2026 is projected at $115 billion to $135 billion, nearly double last year, according to CNBC. That spending is aimed at closing a gap: OpenAI and Anthropic are collectively valued at over $1 trillion, Google’s Gemini has gained traction in the consumer market, and Meta’s last open-source release, Llama 4, failed to captivate developers after a delayed and disappointing debut.
Meta’s stock rose 6.5% on Wednesday, though the rally tracked a broader market jump following President Trump’s announcement of a two-week suspension of Iran attacks.
Meta acknowledged the model has “current performance gaps, specifically long-horizon agentic systems and coding workflows.” For agent builders running local infrastructure on Llama derivatives, the strategic question is whether Meta’s frontier ambitions will eventually produce open-weight models at the Muse Spark capability level, or whether the proprietary turn signals a permanent change in distribution strategy.