Meta Ships Muse Spark, the First Model From Alexandr Wang's Superintelligence Lab

Meta Superintelligence Labs shipped its first model today, April 8, 2026. It’s called Muse Spark, and it’s the first real deliverable from the lab Mark Zuckerberg built around Alexandr Wang after spending $14.3 billion to bring him over from Scale AI last summer. Meta stock jumped roughly 7 to 9% on the announcement, according to Invezz.

The pitch from Meta’s official blog post: Muse Spark is “a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration.” It’s live right now at meta.ai and in the Meta AI app, free to use, with a private API preview opening to select partners.

This is a departure from the Llama playbook in two ways. First, it isn’t open-weights at launch, though Meta says it “hopes to open source future versions.” Second, and more important, this is the first public signal of what the Superintelligence Labs org is actually building, after nearly a year of headlines about nine-figure compensation packages and a restructured research hierarchy.

Three Modes, Not Two

Most current reasoning models give you a fast path and a slow path. Muse Spark adds a third.

Instant runs like a normal chat model. Fast responses for casual queries.
Thinking is extended single-agent reasoning. Same pattern as GPT-5 Thinking or Gemini Deep Think.
Contemplating is the new one. It orchestrates multiple agents in parallel, with each agent running its own reasoning process, and combines their outputs.

Contemplating mode is where Meta is staking its benchmark claims. According to the officechai benchmark breakdown, Muse Spark in Contemplating mode hits 50.2% on Humanity’s Last Exam without tools, compared to 48.4% for Gemini 3.1 Deep Think and 43.9% for GPT-5.4 Pro. On FrontierScience Research it lands at 38.3%, ahead of GPT-5.4 Pro’s 36.7% and well ahead of Gemini Deep Think’s 23.3%.

Meta’s own blog post positions this as efficient test-time reasoning, noting the model implements “thought compression” so that reasoning becomes more efficient after an initial extension phase. In plain terms: the model learns to stop thinking once it has the answer, instead of grinding through more tokens.

The Benchmarks Where It Leads

Thinking mode is where Muse Spark competes with the everyday flagship models from OpenAI, Google, and xAI. Here’s how it stacks up, based on Meta’s published numbers:

Benchmark	Muse Spark	GPT-5.4	Gemini 3.1 Pro
CharXiv Reasoning (figure understanding)	86.4	82.8	80.2
HealthBench Hard	42.8	40.1	20.6
DeepSearchQA (agentic search)	74.8	n/a	69.7
MedXpertQA Multimodal	78.4	77.1	81.3
ZeroBench (visual reasoning, pass@5)	33.0	41.0	29.0

Two categories stand out. HealthBench Hard is a rout: Muse Spark doubles Gemini 3.1 Pro’s score on open-ended health queries and beats Grok 4.2 by more than 20 points. Meta says it worked with “over 1,000 physicians” to curate health training data, and the benchmark numbers back that investment up. The model generates interactive displays explaining things like nutritional content and muscle activation during exercise.

Multimodal understanding is the other win. CharXiv Reasoning tests a model’s ability to parse figures and charts, which is exactly the kind of real-world task where current frontier models still stumble. A 3.6-point lead over GPT-5.4 is meaningful here.

Where It Loses

Meta did not cherry-pick. The company’s own blog acknowledges gaps in “long-horizon agentic systems and coding workflows,” and the published numbers show exactly where:

Benchmark	GPT-5.4	Gemini 3.1 Pro	Muse Spark
ARC-AGI-2 (abstract reasoning)	76.1	76.5	42.5
Terminal-Bench 2.0 (agentic coding)	75.1	68.5	59.0
IPhO 2025 Theory (physics olympiad)	93.5	87.7	82.6
GDPval-AA Elo (office tasks)	1672	n/a	1444

The ARC-AGI-2 gap is the striking one. Gemini 3.1 Pro scores 76.5%. Muse Spark scores 42.5%. That’s a 34-point delta on a benchmark specifically designed to test the kind of novel pattern recognition that separates actual reasoning from memorization. Whatever Meta’s training stack is optimizing for, abstract reasoning isn’t it, at least not yet.

Coding is the other weakness. Terminal-Bench 2.0 measures agentic coding in a real terminal environment, and Muse Spark trails GPT-5.4 by 16 points. For anyone evaluating Muse Spark against GPT-5.4 or Claude Opus 4.6 for day-to-day coding work, the numbers are clear. This isn’t that model. Meta knows it, and said so.

The Pretraining Story

The more interesting claim is buried in the technical section of Meta’s blog post: Muse Spark’s rebuilt pretraining stack reached equivalent capabilities using “over an order of magnitude less compute than Llama 4 Maverick.”

That’s a big deal, and it’s worth reading carefully. Meta isn’t claiming Muse Spark is more capable than Llama 4 Maverick. The claim is that reaching Maverick-level capability required more than 10x less compute with the new stack. If true, this is what Wang’s team was hired to deliver: not a single new model, but a fundamentally more efficient training pipeline that compounds over every future release.

The context here is Llama 4, which Meta released in April 2025 to a muted reception and which triggered the superintelligence reorg in the first place. According to Fortune’s reporting, Meta has committed “hundreds of billions of dollars” to AI infrastructure, and the Wang hire was the leadership piece of that bet. An order-of-magnitude efficiency gain, if it holds up under independent testing, is the kind of return that justifies the spend.

Meta also reports log-linear growth in pass@1 and pass@16 metrics during reinforcement learning, with “gains that generalize predictably to held-out evaluation sets.” That’s research-paper language for: our RL scaling curves are clean and our models aren’t overfitting to their training distributions. Useful, if the numbers hold up.

Multimodal, With Actual Applications

The product demo Meta is leading with isn’t a chart or a benchmark. It’s the photo at the top of this post: you take a picture of a bento box, and Muse Spark estimates the calories of each item, draws labeled dots on the image, and produces a breakdown table. It’s a trivial-sounding demo that’s actually very hard, because it requires entity recognition, localization, visual reasoning, and domain knowledge all stitched together.

Meta’s blog highlights two other use cases:

Interactive minigames. The model can generate playable games inside the chat interface, using visual chain of thought to reason about game state.
Appliance troubleshooting. Point your camera at a broken washing machine, and the model annotates the photo with what to look at and why.

This is Meta playing to its product distribution strength. Google has Lens. Apple has Visual Intelligence. Meta has a billion users across WhatsApp, Instagram, and Facebook, and according to Engadget, Muse Spark will roll out to all of those surfaces in the coming weeks, eventually replacing the Llama models currently powering those chatbots.

Safety and Refusals

Meta also published safety numbers, which is not always standard practice for model launches. The company says Muse Spark “demonstrated strong refusal behavior across high-risk domains including biological and chemical weapons, falling within safe margins on all evaluated frontier risk categories.”

Not much to editorialize on here beyond the fact that they bothered to publish it. Frontier risk disclosures from Meta have historically been thinner than what OpenAI, Anthropic, or Google publish. This one is more in line with industry norms.

What It Means

Three takeaways from reading the announcement and the benchmarks together:

First, Muse Spark is not a GPT-5.4 killer, and Meta didn’t pretend it was. It’s a first release from a brand-new lab, and the model has very real gaps in abstract reasoning and agentic coding. Anyone expecting Meta to leapfrog the frontier labs with its first swing should recalibrate.

Second, the health and multimodal results are genuinely strong. If you’re building anything that touches medical queries, chart parsing, or visual localization, Muse Spark is worth a serious evaluation. HealthBench Hard is not a subtle lead.

Third, the pretraining efficiency claim is the one to watch. If Meta’s stack really trains Maverick-equivalent models with 10x less compute, the next Muse family release will be the real test. Muse Spark is the preview. Muse (whatever comes next) is the thing that will tell us whether the Superintelligence Labs bet is paying off.

The model is free at meta.ai today. Private API preview is open to select partners, with wider availability in the coming weeks. Future versions may be open-sourced, per Meta, though that commitment comes with the usual caveats about capability thresholds and safety review.