Meta has spent the last two years positioning itself as the company keeping AI open. Llama models, released for anyone to download and build on, were the core of that identity. On April 8, 2026, Meta quietly abandoned that position.

Muse Spark, the first model from Meta Superintelligence Labs (MSL), is proprietary. The architecture is not public. The weights are not downloadable. Access is through the Meta AI app, Meta.ai, and a limited private API preview for select developers. The company said it “hopes to open-source future versions,” which is a hedge, not a commitment.

This matters because Meta’s open-source posture was not just branding. It shaped how thousands of businesses thought about AI infrastructure. Llama gave enterprises a route to run capable AI models on their own hardware, under their own control, without paying per-token fees to a US-based AI lab. That option still exists with Llama 4 and older models, but the trajectory has changed.

What Muse Spark Actually Is

The model was built by MSL, the unit Meta assembled after a $14 billion deal that brought in Alexandr Wang from Scale AI as Chief AI Officer. The team’s goal, as publicly stated, is to work toward artificial general intelligence. Muse Spark is their first public output.

Technically, it is described as a “natively multimodal reasoning model” with:

• Multi-agent orchestration — the model can break problems into subagents working in parallel
• Visual chain of thought — it reasons through images, not just text
• Tool use — it can call external tools and APIs as part of completing tasks
• Voice, text, and image inputs — outputs are currently text only

Performance-wise, Meta claims Muse Spark achieves its reasoning capabilities using an order of magnitude less compute than Llama 4 Maverick, its previous mid-size flagship. The implication: a smaller, faster, more efficient model that can do more.

The rollout starts with the Meta AI app and Meta.ai, with expansion to Facebook, Instagram, WhatsApp, Messenger, and Meta’s Ray-Ban AI glasses planned over coming weeks.

The Strategic Shift Worth Noticing

Meta’s move to closed source is not isolated. It is part of a broader pattern across the AI industry: as models become genuinely capable, the companies building them are increasingly reluctant to give away the underlying technology.

The Llama strategy made sense when models were catching up to GPT-4 and Meta needed developer adoption to build ecosystem relevance. Now that frontier models are commercially valuable, keeping the best work proprietary protects the investment. Every lab has arrived at roughly the same conclusion.

For businesses that built AI strategies around the assumption that “open” was the future of enterprise AI, this is a recalibration moment. Open models will continue to exist — the Llama 4 family is still available, as are models from Mistral, Google (Gemma), and others. But the best new models from the best-funded labs are increasingly closed, priced by usage, and optimized for the provider’s commercial interests.

Why Agentic Capability Is the Real Story

The multi-agent orchestration feature in Muse Spark deserves more attention than the open-source debate.

Most AI models today handle tasks sequentially. You give the model a prompt. It generates a response. Complex tasks are broken into separate prompts, one at a time. Muse Spark, according to Meta’s description, can assign parallel subagents to different aspects of the same problem and coordinate their work. This is architecturally closer to how knowledge work actually happens in organizations.

This is meaningful for enterprise AI deployment. Parallel agents that can reason visually, use tools, and coordinate toward a shared outcome can tackle workflows that sequential models cannot handle efficiently. Meta is building these capabilities into the consumer experience first — Meta AI on Instagram and WhatsApp — but the underlying model will be available to developers through the API.

What This Means for Business

The AI vendor landscape is shifting. Businesses that assumed AI was commoditizing toward cheap, open-source models need to revisit that assumption. The frontier is moving to closed, priced-by-usage systems. The choice of which AI providers you build on top of has long-term commercial and strategic implications.

Agentic AI is becoming table stakes. Multi-agent orchestration is no longer an experimental feature. Meta is building it into a model that will reach billions of consumer touchpoints. When your customers are already interacting with parallel AI agents on WhatsApp, the gap between consumer AI and your business operations becomes visible.

Efficiency improvements are real and compound. Muse Spark’s claimed efficiency advantage — similar reasoning at a fraction of the compute — is the kind of improvement that makes AI economically viable for more use cases. Cheaper, faster inference means more opportunities to deploy AI against workflows that were previously too expensive to automate.

The API preview matters if you move fast. Meta is offering private API access to select developers. If Muse Spark’s multi-agent and multimodal capabilities are as described, early API access is worth pursuing before the line gets long.

The broader point is this: AI model development is accelerating, the best models are becoming proprietary again, and agentic capabilities are moving from research into production. For businesses figuring out where to place their AI bets, the window for setting a coherent strategy is narrowing.

If you want to think through what AI agents and AI infrastructure decisions actually mean for your specific business, that conversation is worth having now.

Enterprise DNA put together a free field guide on exactly this: the full Claude ecosystem, Claude Code, and how to roll agents out without breaking things. Get the guide.