Nine months ago, Meta had a problem. Llama 4 had launched to a disappointing reception in April 2025 - widely described as a dud, with Meta later admitting it had used specialized fine-tuned versions to inflate benchmark results. Chinese models from Alibaba's Qwen and DeepSeek were outpacing it on general knowledge and coding. Mark Zuckerberg had already spent $14.3 billion acquiring a 49% stake in Scale AI and installing its co-founder Alexandr Wang as Meta's first-ever chief AI officer, tasked with rebuilding from scratch.
On Wednesday, that rebuilt stack shipped its first product. Meta announced Muse Spark, the first model from Meta Superintelligence Labs - code-named Avocado internally - and the company's first major model release in roughly a year. Meta shares jumped as much as 7% on the news.
What Muse Spark Actually Is
The team rebuilt Meta's entire AI infrastructure from the ground up: new architecture, new data pipelines, new optimization approaches. The result is a natively multimodal model accepting voice, text, and image inputs with text output. It operates in three modes: Instant for casual queries, Thinking for standard reasoning, and Contemplating - a multi-agent parallel reasoning mode designed to compete with Google's Gemini Deep Think and OpenAI's GPT Pro extended thinking. A shopping mode combines the model with Meta's behavioral and interest data, letting it surface personalized purchase recommendations - Meta's clearest near-term monetization angle.
Independent evaluations from Artificial Analysis place Muse Spark fourth on the Intelligence Index v4.0 with a score of 52, behind Gemini 3.1 Pro Preview and GPT-5.4 (both 57) and Claude Opus 4.6 (53). It leads in health and visual understanding. It trails in coding and abstract reasoning - gaps Meta's own technical blog explicitly acknowledged, saying the company continues investing in "long-horizon agentic systems and coding workflows."
A key efficiency claim: Meta says Muse Spark achieves its reasoning capability using more than ten times less compute than Llama 4 Maverick, through a training technique called thought compression that penalizes excessive reasoning tokens during reinforcement learning.
The Open-Source Pivot Nobody Missed
The detail drawing the most scrutiny is not the benchmark table. It is the distribution model. Muse Spark launches as a closed, proprietary model - a direct break from the Llama series that established Meta as the defining force in open-source AI development. The Llama ecosystem reached 1.2 billion downloads by April 2026, averaging one million per day. Self-hosting Llama offered an 88% cost reduction versus proprietary APIs. That ecosystem built Meta enormous developer goodwill.
Wang addressed the pivot directly on X: "Nine months ago we rebuilt our AI stack from scratch. New infrastructure, new architecture, new data pipelines. This is step one. Bigger models are already in development with plans to open-source future versions." Meta has indicated it hopes to release future Muse versions under open-source licenses, and is currently offering private API preview access to select partners.
The developer community is skeptical - and the skepticism is fair. Meta has said similar things before. The more candid reading is that a company spending between $115 billion and $135 billion in capital expenditure in 2026 has decided that open-sourcing its best model architecture while rivals are racing to close a capability gap is a trade it can no longer afford to make.
Muse Spark is available now in the Meta AI app and on meta.ai, with a rollout to Facebook, Instagram, and WhatsApp beginning in the US and expanding internationally in the coming weeks.
