On April 8, Meta unveiled a new AI model called Muse Spark. It is the first model produced by Meta Superintelligence Labs, the research division built around Alexandr Wang, who left his position as CEO of Scale AI to lead a unit with a $14.3 billion mandate and a name that sounds like a Bond villain's subsidiary. [1]

The model does several things at once. It understands images. It uses tools. It performs visual chain-of-thought reasoning, which means it can look at a photograph and work through a problem step by step rather than producing a single-shot answer. It supports multi-agent orchestration, the increasingly popular architectural pattern in which one model coordinates the work of several others. [2] Internally, before it had a public name, it was called Avocado. It was built in nine months. [3]

Nine months is fast. The typical development cycle for a frontier model runs twelve to eighteen months at minimum. But Meta had reasons to move quickly, and those reasons have names. Llama 4 disappointed. The open-weight model that was supposed to demonstrate Meta's competitive position in the AI race instead demonstrated the limits of scaling language models without rethinking the architecture underneath. Investors noticed. The stock dipped. The narrative shifted from "Meta is a serious AI company" to "Meta spends a lot of money on AI."

The spending is not abstract. Meta has guided $115 to $135 billion in AI capital expenditure for 2026 — a number so large it functions more as a statement of intent than a budget line. [4] Muse Spark is the first tangible evidence that this spending is producing something a user can touch. It is rolling out to WhatsApp and Instagram, the two products where Meta's billions of users actually live.

On the Artificial Analysis benchmark — one of the independent evaluations the AI industry uses to compare models — Muse Spark scored 52. [5] That is competitive with frontier models from OpenAI, Anthropic, and Google. It is not category-leading. It does not redefine any benchmark. What it does is demonstrate competence, which, after Llama 4, may be exactly what Meta needed.

The deeper question is strategic. Meta built its AI reputation on open weights — releasing Llama models for anyone to download, modify, and deploy. Muse Spark arrives under a different brand, from a different lab, with a different philosophy. Reuters noted it as the first model from the superintelligence team, a framing that emphasizes the organizational break from Llama's open approach. [4] Whether Muse stays open or goes proprietary will say more about Meta's AI future than any benchmark score.

For now, it works. It is fast. It is in your apps. And somewhere in Menlo Park, an avocado got a promotion.

-- DAVID CHEN, San Francisco