Meta's Muse Spark Is Proprietary — a Sharp Break from Llama's Open-Source Legacy

By Hector Herrera | April 17, 2026

Meta's first major model release under chief AI officer Alexandr Wang and Meta Superintelligence Labs — [Muse Spark,](/business/meta-muse-spark-superintelligence-launch) formerly code-named Avocado — is fully proprietary. The company that spent three years building an open-source AI identity through the Llama family has now shipped a natively multimodal reasoning model it will not be releasing for anyone else to run, fine-tune, or inspect.

This is a strategic reversal, not an incremental adjustment. It signals that Meta's AI ambitions under Wang's leadership have outgrown the business logic that made Llama open-source in the first place.

Context

Meta's Llama series — released starting in 2023 — was one of the most consequential decisions in AI history. By releasing capable models under open or permissive licenses, Meta accelerated the open-source AI ecosystem, drove down the cost of AI experimentation, and created a counterweight to OpenAI and Google's closed approaches. The strategy also served Meta's competitive interests: every developer building on Llama was a developer not paying OpenAI for GPT access, which kept the API market from consolidating.

But the open-source strategy had a monetization problem. Meta's ad-supported business model generates revenue from attention and engagement — not from model access. Llama didn't directly produce revenue. It produced goodwill, ecosystem influence, and AI talent recruitment advantages. Those are real, but they don't appear on a quarterly earnings call.

Alexandr Wang, who joined Meta as chief AI officer in 2025, has a different background: Scale AI, the data labeling and AI evaluation company he founded, was built on proprietary services, enterprise contracts, and recurring revenue. The strategic logic of Muse Spark's closed architecture reflects his fingerprints.

What Muse Spark Is

Architecture: Natively multimodal from the ground up — not a text model with vision bolted on. Muse Spark processes text, images, video, and audio as unified inputs.

Deployment: Rolling out across Meta's apps — Facebook, Instagram, WhatsApp, Messenger — and into Meta's AI glasses hardware. The model is the engine behind the upgraded Meta AI assistant experience.

Capabilities: Meta positions Muse Spark as a reasoning model, placing it in the same competitive category as OpenAI's o-series and Anthropic's extended thinking models. It does not release standard benchmark scores alongside the announcement, which is either a sign of confidence (no need to compete on benchmarks) or a sign of caution (scores aren't competitive enough to lead with).

API access: Not announced. There is no indication that developers will be able to access Muse Spark through an external API in the near term.

Why Meta Closed the Model

Three forces pushed Meta toward a proprietary architecture:

1. Capability concerns. As models become more capable at reasoning, code generation, and multimodal understanding, the risk calculus for open release shifts. A model capable enough to be Meta's primary consumer AI product is capable enough to be misused at scale if weights are public. This mirrors Anthropic's reasoning for not releasing Mythos.

2. Monetization pressure. Meta's AI investment has reached tens of billions annually. Shareholders and Wall Street analysts are pressing the company on AI return on investment. A closed model enables subscription AI products, enterprise licensing, and eventually an external API — revenue streams that an open-weight model cannot support.

3. Competitive intelligence. Open-source models are immediately studied, distilled, and replicated. Chinese AI labs in particular have used Llama weights to accelerate their own development. A proprietary Muse Spark prevents that — which also aligns with the broader industry trend toward protecting frontier capabilities from adversarial distillation (see the Frontier Model Forum intelligence-sharing agreement from the same week).

What Happens to Llama

Meta has not discontinued Llama. The company is expected to continue releasing Llama versions as open or permissive-license models — but as a tier below Muse Spark, not as Meta's primary model. Think of it as the relationship between Google's Gemini and its open-source Gemma family: the open model remains useful for the ecosystem, but the frontier capability stays proprietary.

This creates a two-tier Meta AI strategy: proprietary frontier for apps and eventual monetization, open-weight midrange for developer ecosystem and recruiting.

What to Watch

The critical question is how Meta monetizes Muse Spark beyond its own apps. An external API would generate revenue and compete directly with OpenAI and Anthropic's commercial API businesses — but it would require Meta to build developer trust in a model it has not released for inspection.

The analysts pressing Meta on monetization are right to do so. Llama's value was that it was free and everyone could use it. Muse Spark's value needs to be something different — and Meta hasn't yet articulated what that is.

Source: TechCrunch, April 8, 2026