Meta Launches Muse Spark: A Proprietary Bet on Superintelligence

Nine months after Mark Zuckerberg made a $14.3 billion bet on Scale AI’s Alexandr Wang — the largest talent acquisition in the history of the AI industry — Meta finally has a model to show for it.

On April 8, the company unveiled Muse Spark, the first output of Meta Superintelligence Labs (MSL), the internal organization assembled around Wang after Zuckerberg grew frustrated with the pace and performance of Meta’s existing AI research teams. The model is code-named Avocado and represents a clean break from the strategy that defined Meta’s AI identity for the past three years: open-source.

Muse Spark is proprietary. That single word carries enormous strategic weight.

The End of Llama Supremacy

Meta’s Llama series transformed the AI landscape when it launched in 2023. By releasing capable foundation models under permissive licenses, Meta handed developers worldwide free access to powerful infrastructure, built a vast open-source ecosystem around its software, and positioned itself as the counterweight to closed-source labs like OpenAI and Anthropic.

That strategy ran into trouble in early 2025. Llama 4, released last April, failed to generate the excitement its predecessors had. Developers complained that the models underperformed on benchmarks relative to their scale, lagged behind GPT-5 and Gemini 3 on real-world tasks, and lacked the polish of fine-tuned competitors. The developer community, once an enthusiastic Llama constituency, began routing around Meta’s releases.

Zuckerberg’s response was to dismantle the existing research hierarchy and rebuild. The Wang deal brought not just one person but an entire philosophy: rigorous data engineering, industrial-scale evaluation pipelines, and a willingness to treat AI development as a product discipline rather than a research exercise. Meta Superintelligence Labs was incorporated as a distinct unit with its own compute allocation, hiring authority, and release cadence — insulated from the internal politics that had slowed Llama’s progress.

What Muse Spark Can Do

Meta describes Muse Spark as “small and fast by design, yet capable enough to reason through complex questions in science, math, and health.” The model performs competitively in four primary domains: multimodal perception, structured reasoning, health and wellness applications, and agentic task execution.

The health emphasis is notable. Meta has been quietly building a consumer health AI strategy for two years, piloting features on Instagram and WhatsApp that help users track nutrition, mental health, and fitness. Muse Spark is designed to power this layer at scale — a model that can interpret medical imagery, parse lab results, and engage in nuanced wellness conversations without requiring specialist fine-tuning.

On standard benchmarks, early previews place Muse Spark competitive with GPT-5.4 on reasoning and coding tasks, though Meta has not yet published a comprehensive public evaluation. Independent researchers who received early access described the model’s output quality as “noticeably different from Llama in a good way” — tighter, more consistent, and less prone to hallucination on multi-step problems.

The Rollout

Muse Spark launched immediately as the backbone of the Meta AI app and website, replacing the Llama-based system that had powered those products. Over the following weeks, Meta plans to extend the model to WhatsApp, Instagram, Facebook, Messenger, and its Ray-Ban AI glasses — giving it a potential daily active user base measured in the billions, a distribution advantage no other AI lab comes close to matching.

The glasses integration in particular represents a product direction that Meta has been investing in heavily. Wang’s team reportedly built Muse Spark’s multimodal pipeline with wearable inference as a first-class target, optimizing for low-latency responses over limited connectivity — a technical profile very different from the data-center-first architectures at OpenAI and Google.

Proprietary by Design — For Now

The decision to close-source Muse Spark is framed internally as a competitive necessity. Open-sourcing Llama made sense when Meta needed developer goodwill and was not competing at the frontier. Now that it is — or at least aspires to be — the calculus changes.

That said, Meta has not abandoned open-source entirely. The company signaled that it plans to partially open-source future Muse models, likely releasing smaller distillations or fine-tunable variants for research while keeping the full frontier model proprietary. The move mirrors a pattern seen at other labs: use closed models to capture commercial value, release smaller versions to maintain developer ecosystem loyalty.

For now, the message is clear: Meta is no longer content to be the open-source alternative. It wants to win.

Closing the Gap

Whether Muse Spark delivers on that ambition remains an open question. The model’s debut has been received warmly but cautiously. Reviewers note genuine improvements over Llama 4, but acknowledge that a single model launch does not erase the gap that opened during the past year of Llama’s stagnation.

What matters as much as the model itself is what it signals about Meta’s organizational resolve. Nine months ago, Zuckerberg made an extraordinary commitment of capital and public credibility to this effort. Muse Spark is the first tangible return on that bet. The next model in the Muse series — which Meta has already teased internally — will tell us whether the restructuring produced a lasting step change or a one-time correction.

The AI race is entering a phase where the gap between leaders and followers could widen very quickly. Meta, at least, has made clear it intends to be in the running.