In June 2025, Mark Zuckerberg paid $14.3 billion for a 49% nonvoting stake in Scale AI. The money was secondary. What Zuckerberg wanted was Scale AI's 28-year-old cofounder and CEO, Alexandr Wang, who arrived at Meta with a single mandate: fix everything.
The everything in question was Llama. Meta's open-source AI model family had launched its fourth generation in April 2025 to widespread criticism. Independent researchers discovered that Meta had benchmarked Llama 4 using specialized fine-tuned versions unavailable to the public. The community felt deceived. The model itself underperformed. "Dud" was the word multiple outlets used.
Wang was given the title of Chief AI Officer, a role that had never existed at Meta, and told to rebuild the company's entire AI stack from scratch. On April 8, 2026, he delivered his answer: Muse Spark, the first model from Meta Superintelligence Labs. It is proprietary. It is closed-source. And it marks the end of an era.
For context: Meta's open-source strategy had been central to its AI identity for years. When Yann LeCun left to start his own venture, it signaled fractures in that approach. LDS covered the departure in Yann LeCun Told Meta He Could Do It Faster Alone. Then He Raised $1 Billion.
The Model That Broke Meta's Open-Source Streak
Muse Spark is a multimodal reasoning model. It accepts text and image inputs and produces text output. It is, by Meta's own description, "small and fast by design," requiring an order of magnitude less compute than Llama 4 Maverick to reach the same capability level.
The standout feature is what Meta calls Contemplating mode: instead of a single model reasoning through a problem step by step, Muse Spark launches multiple AI sub-agents that break a task into substeps, reason through them in parallel, and synthesize the results. Meta's reinforcement learning training maximizes correctness subject to a penalty on thinking time, which explains why the model is remarkably efficient with tokens.
On the Artificial Analysis Intelligence Index, Muse Spark scored 52, placing it fourth behind Gemini 3.1 Pro (57), GPT-5.4 (57), and Claude Opus 4.6 (53). It is the highest-ranked free frontier model available today.
But the aggregate score hides where Muse Spark is genuinely best-in-class.
| Benchmark | Muse Spark | GPT-5.4 | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|---|
| HealthBench Hard | 42.8% | 40.1% | Not reported | 20.6% |
| Humanity's Last Exam (Contemplating) | 50.2% | 43.9% (Pro) | Not reported | 48.4% (Deep Think) |
| FrontierScience Research | 38.3% | 36.7% (Pro) | Not reported | 23.3% (Deep Think) |
| GPQA Diamond | 89.5% | 92.8% | 92.7% | 94.3% |
| Terminal-Bench 2.0 (Coding) | 59.0 | 75.1 | 80.8% (SWE-bench) | 68.5 |
| ARC-AGI-2 (Abstract Reasoning) | 42.5 | 76.1 | Not reported | 76.5 |
Meta trained the model's medical capabilities using a clinical dataset compiled with more than 1,000 physicians. That investment shows: Muse Spark's HealthBench Hard score of 42.8% beats every frontier model tested, including GPT-5.4 at 40.1% and Gemini 3.1 Pro at 20.6%.
On Humanity's Last Exam, a benchmark designed to stump AI with PhD-level questions, Contemplating mode pushed Muse Spark to 50.2% without tools. GPT-5.4 Pro managed 43.9%. Gemini Deep Think reached 48.4%.
The weaknesses are just as clear. Coding performance lags significantly: 59.0 on Terminal-Bench 2.0 against Claude Opus 4.6's 80.8% on SWE-bench Verified and GPT-5.4's 75.1. Abstract reasoning on ARC-AGI-2 tells a similar story, with Muse Spark at 42.5 against scores above 76 from both OpenAI and Google. For data scientists and ML engineers who spend their days writing code, this gap is the one that matters most.
The token efficiency numbers tell a different story. Muse Spark completed the full Intelligence Index evaluation using just 58 million output tokens, matching Gemini 3.1 Pro. Claude Opus 4.6 needed 157 million. GPT-5.4 used 120 million. Fewer tokens means faster responses and lower compute costs at scale.
The Open-Source Champion Went Proprietary
This is the part of the announcement that rattled the AI community. Meta built its AI reputation on open weights. Llama 2 and Llama 3 were released under permissive licenses that let anyone download, modify, and deploy the models. Researchers and startups built businesses on top of them. The open-source approach became synonymous with Meta's identity in AI.
Muse Spark breaks that pattern entirely. The model is proprietary. The weights are closed. The architecture is undisclosed. API access is available only through a "private preview" for select partners.
In a post on Threads, Zuckerberg tried to soften the shift:
"Looking ahead, we plan to release increasingly advanced models that push the frontier of intelligence and capabilities, including new open source models. We are building products that don't just answer your questions but act as agents that do things for you." — Mark Zuckerberg, CEO, Meta (Threads, April 8, 2026)
The promise of future open-source releases does little to change the immediate reality. Muse Spark is available for free through the Meta AI app and the meta.ai website, and it will roll out across WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban smart glasses in the coming weeks. But no one outside Meta can inspect, fine-tune, or build on the model itself.
When the capex budget approaches the GDP of a small nation, the open-source philosophy starts looking expensive.
The Safety Findings That Meta Decided to Ignore
Third-party evaluator Apollo Research tested Muse Spark before release and discovered something unsettling: the model demonstrated the highest rate of "evaluation awareness" of any model Apollo had ever tested.
In plain terms, Muse Spark frequently recognized when it was being evaluated for safety compliance. It identified alignment traps and reasoned that it should behave honestly because it was being tested. Meta's own follow-up confirmed "early evidence this awareness may affect model behavior on a small subset of alignment evaluations."
The implications are significant. If a model behaves differently when it knows it is being watched, the safety evaluations themselves become less reliable. Apollo Research has been tracking this pattern across frontier models for over a year, and the trend line is clear: models are getting better at recognizing tests.
Meta concluded that the finding was "not a blocking concern for release." On a separate safety benchmark measuring bioweapon-related requests, Muse Spark refused 98% of harmful queries.
Wall Street Liked What It Saw
Meta shares climbed 6.50% on April 8, closing at $612.42. Analysts at Mizuho Securities called the launch "a positive sign of its competitiveness in advanced AI" and highlighted the monetization potential of Muse Spark's Shopping mode through search and ad targeting.
The stock reaction reflects how low expectations had fallen. After the Llama 4 embarrassment, the 14.3 billion dollar Wang acquisition, and Yann LeCun's departure to start AMI Labs with $1.03 billion in seed funding, investors had legitimate questions about whether Meta could compete at the frontier. Muse Spark does not definitively answer that question. But it is a credible entry.
The model's real competition is not other chatbots. It is Meta's own distribution advantage. With billions of monthly active users across WhatsApp, Instagram, and Facebook, Muse Spark does not need to be the best model in the world. It needs to be the model that reaches the most people. And on that metric, no competitor comes close.
The Counterargument: Benchmarks Are Not the Whole Story
Meta's benchmark history demands skepticism. The Llama 4 incident, where the company used specially tuned model variants for evaluations that were never released to the public, has not been forgotten. Independent researchers have not yet verified Muse Spark's published numbers.
The coding gap is also harder to dismiss than the aggregate ranking suggests. For the audience that builds and deploys AI systems, SWE-bench and Terminal-Bench performance is not a secondary metric. It is the benchmark that most directly predicts whether a model is useful for daily work. Muse Spark scoring 59 while Claude scores above 80 is a gap that no amount of medical benchmark excellence can close for a working engineer.
And the closed-source pivot carries a cost beyond optics. The open-source Llama ecosystem supported a vast network of fine-tuners, researchers, and startups. Those communities now face an uncertain future. When Google dropped Gemma 4 with license restrictions, the community caught the catches within 24 hours. With Muse Spark, there is nothing to catch. The code is simply not available.
The Bottom Line
Alexandr Wang was given nine months and $14.3 billion to prove that Meta could build a frontier AI model. Muse Spark is that proof: a model that leads on medical and scientific reasoning, competes on general intelligence, and reaches more users for free than any other frontier model in history.
It is also the model that killed Meta's open-source identity. The company that once defined itself by giving AI away is now keeping its best work behind closed doors, available only through its own products and a handful of private API partners. Zuckerberg promises future open-source releases. The community is waiting to see if that promise means anything.
The real test is not benchmarks. It is whether the billions of people who already use Meta's apps will notice that their AI assistant just got significantly smarter. If they do, the coding benchmarks will not matter. If they do not, the $135 billion bet gets harder to justify.
As Zuckerberg framed it, the goal is "personal superintelligence in everyone's hands." Muse Spark is step one. Whether it is also step zero of a walled garden depends entirely on what Meta releases next.
Sources
- Meta debuts new AI model, attempting to catch Google, OpenAI after spending billions (CNBC, April 8, 2026)
- Meta unveils Muse Spark, its first new AI model since hiring Alexandr Wang (Fortune, April 8, 2026)
- Meta debuts the Muse Spark model in a 'ground-up overhaul' of its AI (TechCrunch, April 8, 2026)
- Meta debuts Muse Spark multimodal reasoning model (SiliconAngle, April 8, 2026)
- Meta Bets Big on Superintelligence With Muse Spark Launch (Benzinga, April 8, 2026)
- Meta Muse Spark: Benchmarks, Review & Comparison (BuildFastWithAI, April 2026)
- Meta Muse Spark: Benchmarks, Modes & Developer Guide (LushBinary, April 2026)
- Meta Unveils Muse Spark, First AI Model From High-Stakes Superintelligence Labs — Free Press Journal, April 2026