Meta just dropped Muse Spark — its first frontier AI model built from scratch — and it signals something bigger than a product launch. Meta Muse Spark is the first model from Meta Superintelligence Labs, the unit led by Alexandr Wang after Meta’s $14.3 billion Scale AI deal. And here’s the part that matters: it’s closed-source. The company that spent years evangelizing open-weight AI with Llama just shipped a proprietary model. That strategic reversal tells you more about where the AI industry is heading than the benchmarks do.
Why Meta Muse Spark Changes Meta’s AI Playbook
For three years, Meta was the loudest voice in the open-source AI movement. Llama 2 and Llama 3 became the backbone of thousands of startups, enterprise deployments, and research projects. Mark Zuckerberg personally argued that open models were the future.
Then Llama 4 underperformed. The Maverick and Scout variants launched to mixed reviews. Competitors pulled ahead. And Meta made a calculation: competing at the frontier requires keeping your best work proprietary — at least temporarily.
Muse Spark, originally codenamed “Avocado,” was built from the ground up over nine months by Meta Superintelligence Labs. It’s not a Llama derivative. It’s a new architecture, a new training pipeline, and a new strategy. Meta says it hopes to open-source future versions, but the message is clear: the competitive edge comes first.
This isn’t unique to Meta. Every major lab is now running a dual-track strategy — open models for ecosystem growth, closed models for competitive positioning. But Meta saying it out loud, after being the most visible open-source advocate in the industry, is a watershed moment.
The Alexandr Wang Factor
The person driving this shift is Alexandr Wang, who became Meta’s first-ever Chief AI Officer in June 2025. The backstory matters.
Meta acquired a 49% nonvoting stake in Scale AI for $14.3 billion — one of the largest AI deals in history. Wang stepped down as Scale’s CEO (he’d founded the company at 19) and moved to Meta to lead a new division: Meta Superintelligence Labs.
Wang’s background is data infrastructure and model evaluation, not model architecture. That perspective shows in Muse Spark’s design priorities. The model emphasizes efficiency and multimodal perception over raw parameter counts. Meta claims Muse Spark achieves its reasoning capability using more than 10x less compute than Llama 4 Maverick, driven by a training technique called “thought compression.”
Nine months from Wang’s arrival to a frontier model launch is fast by any standard. It suggests Meta’s problem wasn’t talent or compute — it was organizational focus. Wang consolidated the effort under one lab with one mandate: build the best model possible, open-source strategy be damned.
Benchmark Reality: Where Muse Spark Wins and Loses
Let’s look at the actual numbers. Muse Spark scores 52 on the Artificial Analysis Intelligence Index v4.0, placing it fourth among all models benchmarked:
| Model | Intelligence Index Score |
|---|---|
| Gemini 3.1 Pro Preview | 57 |
| GPT-5.4 | 57 |
| Claude Opus 4.6 | 53 |
| Muse Spark | 52 |
Fourth place on a first attempt from a new lab is genuinely impressive. But the category-level performance tells a more interesting story.
Where Muse Spark Leads
Healthcare and medical reasoning. Muse Spark scores 42.8 on HealthBench Hard, beating GPT-5.4 (40.1) and crushing Gemini 3.1 Pro (20.6). This is the single benchmark where Muse Spark definitively outperforms every frontier model. For healthcare enterprises evaluating AI, this is a significant data point.
Multimodal vision understanding. An 80.5% score on MMMU-Pro makes Muse Spark the second-most capable multimodal model, behind only Gemini 3.1 Pro Preview (82.4%). Meta’s investment in image understanding — driven by Instagram and Facebook’s visual nature — is paying off.
Token efficiency. Muse Spark used just 58 million output tokens to complete the full Intelligence Index evaluation, comparable to Gemini 3.1 Pro (57M) and dramatically lower than Claude Opus 4.6 (157M) or GPT-5.4 (120M). In production, this means lower inference costs.
Where Muse Spark Falls Short
Coding. A Terminal-Bench 2.0 score of 59.0 trails GPT-5.4 (75.1) and Gemini 3.1 Pro (68.5) by a wide margin. Meta acknowledges this gap directly. For developers evaluating coding assistants, Muse Spark isn’t competitive yet.
Abstract reasoning. Muse Spark scores 42.5 on ARC-AGI-2, while GPT-5.4 (76.1) and Gemini 3.1 Pro (76.5) score nearly double. This is the widest gap in the benchmark suite.
Agentic tasks. On GDPval-AA (real desktop and office task performance), Muse Spark’s 1,444 ELO trails GPT-5.4 (1,674) by 230 points and Claude Opus 4.6 (1,607) by 163 points. In the age of AI agents, this gap matters.
The “Contemplating” Mode: Meta’s Answer to Reasoning Models
Muse Spark introduces a feature Meta calls “Contemplating” — a reasoning mode that runs sub-agents in parallel before responding. This is Meta’s answer to OpenAI’s chain-of-thought reasoning and Anthropic’s extended thinking.
The implementation is different from competitors. Rather than a single chain of thought, Contemplating spins up multiple reasoning paths simultaneously and synthesizes results. Meta claims this produces more reliable outputs on complex queries while keeping latency manageable.
This parallel reasoning approach aligns with the “thought compression” training technique — the model learns to reach conclusions through compressed reasoning paths rather than long sequential chains. It’s an interesting architectural bet, and the healthcare benchmark results suggest it works well for domains requiring careful analysis.
What This Means for the AI Landscape
The End of the Open-Source Consensus
Meta going closed-source on its flagship model is a signal the rest of the industry should pay attention to. Llama will continue — Meta confirmed ongoing development of open-weight models. But the frontier work happens behind closed doors now.
For startups and enterprises that built on Llama, the message is nuanced. You can still use open-source AI models for production workloads. Gemma 4 with Apache 2.0 licensing and Llama 4 Maverick remain excellent options. But the best Meta model? That’s proprietary, and the API is invite-only.
Healthcare AI Gets a New Contender
Muse Spark’s HealthBench dominance is the most strategically interesting benchmark result. Healthcare AI is a massive market — estimated at $45 billion by 2027 — and Meta just demonstrated the strongest medical reasoning capability in any general-purpose model.
Meta hasn’t announced specific healthcare partnerships, but the benchmark performance gives it a credible pitch to health systems and pharmaceutical companies. Combined with Meta’s existing relationships through WhatsApp (used by health systems globally), this could be a significant enterprise play.
The Four-Way Frontier Race
The AI industry now has four genuine frontier competitors:
- OpenAI — GPT-5.4 leads on coding and agentic tasks, with GPT-5.5 “Spud” reportedly weeks away
- Google — Gemini 3.1 Pro ties for the top Intelligence Index score and leads on abstract reasoning
- Anthropic — Claude Opus 4.6 dominates real-world work tasks (GDPval) and extended context applications
- Meta — Muse Spark brings healthcare leadership, efficiency, and 3 billion monthly active users as a distribution moat
That last point is critical. OpenAI has ChatGPT. Google has Search. Anthropic has enterprise contracts. But Meta has the largest consumer distribution platform on Earth. Every Facebook, Instagram, and WhatsApp user is a potential Muse Spark user. The model doesn’t need to win every benchmark — it needs to be good enough for 3 billion people.
How to Access Muse Spark Today
Muse Spark is available now through two channels:
- Meta AI app and meta.ai website — free, available immediately
- Private API preview — select partners only, paid API access planned for later
The model will roll out to Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses in the coming weeks. All consumer-facing flavors are free to use, though Meta may impose rate limits.
For enterprise buyers evaluating Muse Spark, the current limitation is clear: no public API means no production integration yet. If your use case is healthcare or multimodal understanding, it’s worth joining the API waitlist. For coding or agentic workflows, GPT-5.4 and Claude Opus 4.6 remain the better choices.
What Comes Next
Three things to watch:
GPT-5.5 timing. OpenAI’s “Spud” completed pretraining in late March. Sam Altman says it’s weeks away. If GPT-5.5 launches in April, it could overshadow Muse Spark before Meta even finishes its rollout.
Meta’s API pricing. When the paid API opens, pricing relative to OpenAI and Anthropic will determine whether Muse Spark becomes an enterprise option or stays a consumer product.
Open-source follow-through. Meta says it hopes to open-source future Muse models. If that happens, the open-source AI ecosystem gets a frontier-quality model with a new architecture. If it doesn’t, Meta loses the community goodwill that made Llama successful.
The bottom line: Meta just proved it can build a frontier model outside the Llama lineage, in nine months, with competitive benchmarks and genuine category leadership in healthcare AI. The closed-source decision is controversial but rational. And with 3 billion users as distribution, Muse Spark doesn’t need to be the best model — it just needs to be the most accessible one.
FAQ
What is Meta Muse Spark?
Muse Spark is Meta’s newest AI model, the first built by Meta Superintelligence Labs under Chief AI Officer Alexandr Wang. It’s a multimodal model that accepts text, voice, and image inputs, features a “Contemplating” reasoning mode, and is currently available through the Meta AI app and meta.ai website.
Is Meta Muse Spark open-source like Llama?
No. In a significant departure from Meta’s open-source AI strategy, Muse Spark is a closed, proprietary model. Meta has stated it hopes to open-source future versions of the Muse model family, but the current model is only available through Meta’s own products and an invite-only API preview.
How does Muse Spark compare to GPT-5.4 and Claude?
Muse Spark ranks fourth on the Artificial Analysis Intelligence Index (score: 52), behind Gemini 3.1 Pro (57), GPT-5.4 (57), and Claude Opus 4.6 (53). It leads all models on healthcare benchmarks (HealthBench Hard: 42.8) and multimodal vision tasks, but trails significantly on coding, abstract reasoning, and agentic task performance.
Why did Meta hire Alexandr Wang from Scale AI?
Meta acquired a 49% stake in Scale AI for $14.3 billion in June 2025 and brought in founder Alexandr Wang as its first Chief AI Officer. Wang’s expertise in data infrastructure and model evaluation was seen as critical to accelerating Meta’s frontier AI development after Llama 4 underperformed against competitors.
Can developers use the Muse Spark API?
Not yet publicly. A private API preview is available to select partners, with paid API access planned for a broader audience at a later date. Consumer access through the Meta AI app and social platforms is free, with potential rate limits.