On April 8, 2026, Meta announced Muse Spark — its first major model in over a year, and the debut output from Meta Superintelligence Labs (MSL). The model rolled out immediately across Facebook, Instagram, WhatsApp, Messenger, and Meta’s Ray-Ban smart glasses, reaching a combined user base of more than 3 billion people. It competes, at least on paper, with GPT-5.4 and Gemini 3.1 Pro.

And it’s closed source.

That last detail matters more than the benchmarks. Meta built its reputation in AI on Llama — a series of open-weight models that developers could download, fine-tune, and deploy freely. Muse Spark ends that era. According to Bloomberg, the model’s design and code won’t be made public. Meta now treats its most capable AI as a commercial asset.

Quick Summary: What Just Happened

Detail	Info
Date	April 8, 2026
Model	Muse Spark (internal codename: “Avocado”)
Built by	Meta Superintelligence Labs, led by Alexandr Wang
Distribution	Facebook, Instagram, WhatsApp, Messenger, Ray-Ban glasses
License	Closed source — breaks from Llama open-weights tradition
Benchmark rank	4th on Artificial Analysis Intelligence Index v4.0 (score: 52)
Official Source	about.fb.com announcement

Bottom line: Meta’s biggest AI move in a year is also a quiet reversal of everything Llama stood for. The community gets nothing to download.

What Actually Happened

This wasn’t a Llama update. According to TechCrunch, Muse Spark is a “ground-up overhaul” — a full rebuild of architecture, infrastructure, and data pipelines. The team achieved comparable capability to Llama 4 Maverick using over 10x less compute. This isn’t a revision. It’s a departure.

The model accepts voice, text, and image inputs. Output is text-only in the initial release. It currently powers Meta AI on meta.ai and the standalone Meta AI app, with broader deployment to Facebook, Instagram, WhatsApp, Messenger, and Meta’s Ray-Ban smart glasses underway. Meta says it’s currently free, with rate limits possible. A paid API is planned after an initial private preview period.

Per CNBC, the launch follows Meta’s $14.3 billion move in June 2025 to bring Alexandr Wang aboard as Chief AI Officer — acquiring a 49% nonvoting stake in Wang’s company Scale AI in the process. Muse Spark is his first major product output, built over approximately nine months by the team he assembled.

The Llama Pivot: Why This Is a Bigger Deal Than Any Benchmark

Meta has released Llama models openly since 2023. Llama 2, Llama 3, Llama 4 — all available for download, fine-tuning, local deployment, commercial use. That openness wasn’t just PR. It was genuine community infrastructure. Thousands of companies, researchers, and individuals built on Llama weights.

Muse Spark changes the relationship.

VentureBeat characterized the move as a significant strategic reversal from Meta’s open-weights tradition. Meta told reporters it “hopes to open-source future versions of Muse” — language that reads more like a hedge than a commitment. For now, there are no weights to download, no architecture paper, no reproducibility. Just an API.

Compare that to how Meta previously talked about Llama. Mark Zuckerberg made open-source AI a philosophical position, not just a product decision. He argued in 2024 that open-source AI was safer and better for the world. That position is now, quietly, in the past.

Why the reversal? The economic answer isn’t subtle. Meta’s AI capex plan for 2026 sits at $115–135 billion. You don’t spend that and give the output away for free. Closed-source is how you capture the value.

What Muse Spark Can Actually Do

The benchmarks are real, but they need context.

Muse Spark scores 52 on the Artificial Analysis Intelligence Index v4.0, placing it fourth overall. GPT-5.4 and Gemini 3.1 Pro both score 57. Claude Opus 4.6 sits at 53. Muse Spark is competitive — genuinely in the tier — but the claim that it challenges the top models isn’t quite right across the board. Meta’s framing is accurate. Careful, but accurate.

Where it does lead: HealthBench Hard, a medical AI benchmark, where Muse Spark scores 42.8 — above the competition. If you’re building in health tech, that’s not a small detail. Meta has positioned health AI as one of its flagship use cases, and Muse Spark appears to back that up.

Beyond benchmarks, the model covers multimodal input (voice, text, image), reasoning, and agentic tasks. No output modalities beyond text in the initial release, though deployment across audio-capable Ray-Ban hardware suggests that changes. For a fuller picture of how this generation of models stacks up, see our 2026 AI models comparison.

The Largest Single AI Model Deployment in History

Here’s the number that reframes everything else: 3 billion+. That’s the approximate combined user base of Facebook, Instagram, WhatsApp, and Messenger. Meta is deploying Muse Spark across all of them at once.

No model has ever shipped to this many potential users on day one. Not GPT-4. Not Gemini Ultra. Not Claude. The distribution math doesn’t compare.

The Ray-Ban smart glasses are a smaller number but a different signal. Putting a closed-source proprietary model into wearable hardware is a deliberate choice. Every voice query, every ambient interaction on those glasses now runs on infrastructure Meta owns and controls entirely. That’s not just a product decision — it’s a data strategy.

Ambient, persistent AI running across 3 billion devices is exactly what agentic AI advocates have been pointing toward. Our breakdown of the agentic AI shift from GTC 2026 covers why ambient, persistent AI is where the whole industry is heading.

What Developers Who Built on Llama Need to Know

Existing Llama models aren’t going anywhere. Llama 4 still exists. The weights are still downloadable. Nothing about the Muse Spark launch retroactively changes what Llama offers.

But the trajectory has shifted.

Meta’s most capable, most actively developed model is now proprietary. Future flagship releases will likely follow the same path. The open Llama ecosystem may continue — there’s no indication Meta is abandoning it — but it’s no longer where Meta’s best work goes.

For teams evaluating open-weight versus API-based models, this changes the calculus. Llama was a serious contender for local deployment precisely because it kept pace with commercial models. If Meta’s frontier work stays closed, that parity erodes over time.

The current alternatives for open-weight deployments look increasingly toward Chinese labs. See our analysis of Seedance, Doubao, Qwen, and DeepSeek for the current state of that landscape, or our DeepSeek V4 review for the leading open-weight option right now.

How We Got Here: The Alexandr Wang Factor

Wang joining Meta in June 2025 was the tell.

Scale AI built its business on data labeling and AI infrastructure — the essential, unglamorous work that makes frontier models possible. Wang wasn’t just a talent acquisition. He was a statement about what Meta wanted to build: competitive frontier models, not contributions to the commons.

Meta Superintelligence Labs was the organizational structure that followed. A separate unit, distinct from Meta AI Research. Focused on building frontier models rather than publishing papers or releasing open weights. The name itself is a signal. “Superintelligence Labs” is not what you call a team whose primary output is Llama releases.

Muse Spark is what that team built.

The Bigger Picture

Meta’s pivot isn’t happening in isolation. It’s a pattern.

OpenAI started as a nonprofit committed to safe, open AI. It’s now one of the most valuable closed-source AI companies on earth. Google published transformer research that others built on. It now keeps its best models proprietary. The entire frontier AI industry has converged on the same economic conclusion: open weights are a community good that don’t pay for $100 billion in compute.

Meta held out longer than most. The Llama releases were genuine. The developer goodwill was earned. Muse Spark is the moment Meta officially joined the rest of the industry.

For users of Meta’s platforms, this may not matter much. Meta AI gets smarter. The experience improves. The model powering it is closed — but so is the model powering Siri, Alexa, and every other consumer AI assistant people have been using for years.

For developers who built on the assumption that Meta’s trajectory meant increasingly capable open weights, this is the clearest possible signal: that assumption no longer holds.

Our Take

The benchmarks are fine. Muse Spark is a real, competitive model. If health AI is your domain, it leads the field. Across the broader evaluation suite, it’s a credible fourth — better than most companies will ever ship.

But the story isn’t the benchmarks. The story is the strategy.

Meta built real political capital in the AI developer community with Llama. Researchers ran it on laptops. Startups built entire products on it. Companies in regulated industries used it specifically because local deployment meant no data leaving the building. That value was real, and Meta knew it.

Muse Spark closes that chapter. The pivot is rational given the capital commitment — $115–135 billion in AI infrastructure is not a bet you make and then give away the returns. But rational doesn’t mean costless. There’s a community that oriented around Meta’s openness, and that community is now an API customer.

Meta says it hopes to open-source future Muse versions. Hold that lightly. The incentives now point in one direction, and they’re not pointing toward free weights.

---

Frequently Asked Questions

What is Meta Muse Spark?

Muse Spark is Meta’s first model from Meta Superintelligence Labs, released April 8, 2026. It’s a closed-source, multimodal model — accepting voice, text, and image inputs, with text output — deployed across Facebook, Instagram, WhatsApp, Messenger, and Meta’s Ray-Ban smart glasses. On the Artificial Analysis Intelligence Index v4.0, it scores 52, placing it fourth behind GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6.

Why is Muse Spark different from previous Meta AI models?

Previous Meta flagship models — including the Llama series — were released as open-weight models that anyone could download, fine-tune, and deploy. Muse Spark is closed source. Its architecture, training code, and weights are not publicly available. This is a deliberate shift in Meta’s strategy, treating frontier AI as commercial IP rather than community infrastructure.

Who built Muse Spark?

Meta Superintelligence Labs, a unit formed under Chief AI Officer Alexandr Wang. Wang joined Meta in June 2025 after Meta acquired a 49% nonvoting stake in his company Scale AI for approximately $14.3 billion. Muse Spark (internally codenamed “Avocado”) was built over roughly nine months.

How does Muse Spark compare to GPT-5.4 and Gemini 3.1 Pro?

Muse Spark scores 52 on the Artificial Analysis Intelligence Index v4.0 — fourth overall. GPT-5.4 and Gemini 3.1 Pro both score 57; Claude Opus 4.6 scores 53. Muse Spark outperforms all competitors on HealthBench Hard (score: 42.8). It’s genuinely competitive but does not lead across multimodal, reasoning, and agentic benchmarks overall.

Can developers access Muse Spark via API?

Not publicly yet. Meta announced a private API preview is coming, with a paid API planned following that period. Currently, Muse Spark is accessible through Meta AI on meta.ai and the Meta AI app, with rollout to WhatsApp, Instagram, Facebook, and Messenger underway.

Does this mean Llama is going away?

No. Existing Llama models remain available for download and use. But Meta’s flagship development has moved to the proprietary Muse series. The Llama ecosystem continues — there’s no announced end-of-life — but it no longer represents where Meta’s best capabilities are going.

Is Meta planning to open-source future Muse models?

Meta stated it “hopes to open-source future versions” of Muse Spark, but gave no timeline and released no architecture details alongside the launch. Given the commercial context — $115–135 billion in AI capex for 2026 — treating the Muse series like Llama would run against the current incentive structure.

---

Last updated: April 9, 2026. Sources: Meta announcement, TechCrunch, Bloomberg, CNBC, CNBC follow-up, VentureBeat, Artificial Analysis Intelligence Index, HealthBench Hard. Related reading: AI Models Compared 2026, Agentic AI Is the New Default, DeepSeek V4 Review, China AI Models Compared