Meta Launches Muse Spark — Its First Closed-Source AI Model and a Break From the Llama Playbook
Meta Superintelligence Labs ships Muse Spark, a natively multimodal reasoning model that marks the company's first closed-source AI release. The move signals a major strategic shift away from the open-weights Llama family.
Meta shipped Muse Spark on April 8 — the first model out of Meta Superintelligence Labs, the unit formed around Scale AI founder Alexandr Wang. No open weights. No fine-tuning access. A full proprietary deployment. That is a significant break from the Llama strategy that made Meta the dominant force in open-source AI over the past two years.
What Muse Spark Can Do
The model is natively multimodal from day one. It accepts voice, text, and image inputs in any combination — not bolted on after the fact, as was the case with earlier Llama releases. Meta’s headline differentiator is “Contemplating mode,” a parallel multi-agent reasoning system that spawns multiple internal reasoning threads on complex tasks and synthesizes the best response. Think chain-of-thought, but parallelized at inference time.
Muse Spark is already live in the Meta AI app, which shot to #5 on the US App Store within 24 hours of the announcement. Rollout to Facebook, Instagram, WhatsApp, Messenger, and Meta Ray-Ban glasses is expected over the coming weeks.
Why Meta Is Going Closed
The pivot away from open weights is deliberate. Meta’s previous open-source releases — Llama 2, Llama 3, and Llama 4 Scout — handed competitors free baseline models that closed the gap on Meta’s own products. With Muse Spark, the company is betting on a proprietary moat. Alexandr Wang’s influence is visible: Scale AI’s business was built on the premise that data quality and RLHF pipelines matter more than raw model size, and Muse Spark is the first Meta model that publicly puts that theory at the center of its training stack.
Meta has not disclosed parameter count, context window, or API pricing. What it has confirmed: Muse Spark will power the unified Meta AI across all major surfaces — a very different deployment model than anything the Llama family was designed for.
What This Means for the Open-Source Community
The Llama family is not going anywhere. Meta has confirmed that Avocado and Mango — two upcoming models under separate codenames — will still ship with open weights. But Muse Spark’s proprietary release makes clear that Meta is now running a two-track strategy: open models for developer adoption and ecosystem growth, closed models for consumer product competitiveness.
For developers who have built on Llama, nothing changes today. For anyone evaluating frontier models for new projects, Muse Spark is a serious new entrant — though API access details and benchmark disclosures will determine whether it can actually compete with GPT-5.4, Claude Opus, or Gemini 3 on the tasks that matter most.
The clearest signal here is what it says about the economics of open-source AI. When even Meta — the company that positioned open weights as a core strategic differentiator — ships its most capable model as proprietary, the argument that openness is always the right default gets harder to sustain. The frontier is expensive. Closed models pay for themselves. Open models need a different revenue story.