Meta Muse Spark: The End of Open Source AI at Meta?
Meta launches Muse Spark, its first proprietary model. After years of championing Llama and open source, why the sudden pivot? Breaking down the shift that's reshaping AI in 2026.
$14.3 billion. That’s how much Meta paid to recruit Alexandr Wang and invest in Scale AI back in June 2025. Nine months later, the first result of that massive bet has landed: Muse Spark, a proprietary AI model. Yes, proprietary — from Meta, the company that turned open source AI into its brand identity with Llama.
On April 8, 2026, Meta unveiled Muse Spark, the first model built by its Meta Superintelligence Labs. The contrast with the company’s historical strategy is striking. While Zhipu AI was publishing GLM-5.1 under the MIT license the day before, Meta was locking the doors on its newest creation.
This isn’t just a product launch. It’s a strategic signal that could reshape the AI landscape for the rest of 2026.
Meta and Open Source AI: A Love Story That’s Falling Apart
To understand why Muse Spark matters, you need the backstory. Since 2023, Meta has been the undisputed champion of open source AI. Llama became as recognizable as GPT in AI circles. The strategy was clear and openly championed by Mark Zuckerberg himself: release model weights for free, build an ecosystem, attract developers, and prevent OpenAI and Google from locking down the market.
The economic logic was bulletproof. By giving Llama away, Meta ensured that thousands of companies and developers built on its ecosystem. Every startup that fine-tuned Llama became an unwitting ambassador. Every research paper that used Llama reinforced Meta’s AI legitimacy.
And it worked. Llama 2 exploded in 2023, racking up millions of downloads. Llama 3 dominated the open source space in 2024, becoming the go-to model for on-premise deployments. The community loved Meta — a Big Tech giant actually playing the transparency game in a sector that was increasingly closing ranks.
And then came Llama 4.
In April 2025, Meta launched Llama 4, and that’s when everything fell apart. The model was “widely panned as a dud,” as Fortune put it. But the real issue wasn’t performance — it was the cheating.
Meta got caught red-handed manipulating benchmarks. The company was using specialized, fine-tuned versions for specific tasks — versions unavailable to the public — to inflate scores. The version developers could actually download didn’t perform at the same level. Meta eventually admitted it publicly.
For a company that had built its credibility on transparency, this was devastating. The open source community, accustomed to trusting Meta’s promises, felt betrayed. Hacker News and Reddit discussions were brutal. Developers migrated to other solutions: Claude Code for coding, Chinese open source models for everything else.
Behind closed doors, Zuckerberg drew the inevitable conclusion: open source alone was no longer enough to stay in the race.
Alexandr Wang and the $14 Billion That Changed Everything
In June 2025, Meta made a move nobody saw coming. The company invested $14.3 billion in Scale AI and recruited its founder, Alexandr Wang, to lead a new division: the Meta Superintelligence Labs (MSL).
Wang is no ordinary hire. At 27 when he was recruited, he had already built Scale AI into the company that supplied training data to every major AI lab — OpenAI, Google, Anthropic. He knew the internals of every frontier model.
His mandate at Meta was explicit: catch up with OpenAI, Anthropic, and Google, whatever it takes. And his vision translated into a radical shift.
“Over the past nine months, Meta Superintelligence Labs has rebuilt our AI stack from the ground up, moving faster than any development cycle we’ve ever run,” Meta stated in its launch blog post.
The result? Muse Spark. The internal code name was “Avocado,” revealed by CNBC as early as December 2025. And unlike anything Meta had built before, this model is not open source.
Muse Spark: What the Model Does (and Doesn’t Do)
Let’s be precise about what Muse Spark actually is. Meta isn’t positioning it as the best model on the market — and that’s a deliberate choice.
Capabilities
- Reasoning: Meta’s first step-by-step reasoning model, with a “contemplation” mode that launches sub-agents in parallel
- Native multimodal: text and images, both input and output
- Multi-agent orchestration: can coordinate multiple sub-agents for complex tasks
- Tool use: built-in external tool integration
Benchmarks
Meta published results that put Muse Spark in the race — without dominating:
| Benchmark | Muse Spark | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| GPQA Diamond (PhD reasoning) | 89.5% | 92.7% | 92.8% | 94.3% |
| HealthBench Hard (healthcare) | 42.8% | — | 42.1% | — |
The model excels in healthcare and scientific reasoning but lags behind in coding and long-running agentic tasks — precisely the domains where Claude and GPT dominate.
Meta openly acknowledges this: “We continue to invest in areas where performance still falls short, particularly long-running agentic systems and developer workflows.”
The “contemplation” mode is technically interesting. Instead of chaining reasoning tokens in a single stream (like Claude and GPT do with their “thinking” modes), Muse Spark can launch multiple sub-agents in parallel, each exploring a different facet of the problem. Meta claims this approach allows it to “rival the extreme reasoning modes of frontier models like Gemini Deep Think and GPT Pro.”
Access: More Locked Down Than Ever
Here’s the crucial point. Muse Spark is:
- Open source ❌
- Open weight ❌
- Available via public API ❌
The model is available exclusively through the Meta AI app and meta.ai. Rollout to WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban glasses is planned for the coming weeks. A handful of “selected partners” have access to a private API preview. That’s it.
In short: Muse Spark is even more proprietary than its competitors’ models. At OpenAI, you can buy API access for a few dollars per million tokens. At Anthropic, same deal. At Meta? You use their app or nothing at all.
The irony is almost comical. For two years, Zuckerberg criticized OpenAI for being “a nonprofit that acts like a closed company.” And now Meta launches a model that’s even more locked down than GPT.
Why Meta Abandoned Open Source (For Now)
The obvious question: why? Meta had gone all-in on openness. What changed?
1. The Llama 4 Failure Broke the Business Model
Open source worked as long as Meta could say: “Our models are as good as the proprietary ones, and they’re free.” When Llama 4 turned out to be disappointing and the benchmark scandal broke, that value proposition collapsed.
Developers migrated to Claude, GPT, and Chinese open source models. Meta lost the mindshare it had built.
2. The Race to Superintelligence Doesn’t Mix With Openness
The division’s name says it all — Superintelligence Labs. Wang’s goal isn’t to build a good open source model. It’s to build the best model in the world. And in that race, publishing your weights is like showing your hand to every competitor at the table.
OpenAI and Anthropic figured this out long ago. Meta is finally admitting the same thing.
3. Capex Is Exploding — ROI Is Non-Negotiable
Meta plans to spend between $115 and $135 billion on AI infrastructure in 2026 — nearly double its 2025 budget. At that level of investment, giving models away for free is no longer sustainable. The global generative AI market is projected to reach $325 billion by 2033, according to Grand View Research. Meta wants its slice.
Hence the “private API preview” and plans for a paid API — exactly the business model Zuckerberg was criticizing at OpenAI and Anthropic two years ago. Meta stock jumped 6.5% on the announcement day, a clear signal that Wall Street approves the shift toward direct monetization.
4. Geopolitics Has Changed the Math
While Meta was locking down its model, Zhipu AI (Z.ai) published GLM-5.1 under the MIT license. The Chinese model beat GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro with a score of 58.4 — and it’s completely free. The absolute paradox: China is opening up while the United States closes down.
This geopolitical reversal makes Meta’s strategy even more jarring. The company that criticized OpenAI for being closed is now adopting an even more restrictive model, while Chinese labs pick up the open source torch.
What Muse Spark Actually Means for You
If You’re a Developer
The bad news: you can’t fine-tune Muse Spark or deploy it on your own servers. The model is a black box embedded in Meta’s ecosystem.
The good news: open source AI isn’t dead. It’s just moving. Here are the open weight alternatives available today:
| Model | Developer | Params (total/active) | License | Specialty |
|---|---|---|---|---|
| GLM-5.1 | Zhipu AI | 744B / 40B | MIT | Agentic coding (8h autonomous) |
| Gemma 4 | 27B (dense) | Apache 2.0 | Multimodal (text+image+audio) | |
| Qwen 3.6-Plus | Alibaba | MoE | Open | Agents, 1M context |
| Mistral Small 4 | Mistral | 119B MoE | Apache 2.0 | Efficiency, EU deployment |
| gpt-oss-120b | OpenAI | 120B | Open weight | OpenAI’s first open model |
As of April 2026, 5 of the 6 major open source model families use MoE architecture and ship under permissive licenses. The landscape has never been richer — with or without Meta.
If You’re a Business
Meta’s pivot sends a clear signal: the era of tech giants giving away their best models for free is ending. Frontier models are becoming paid products. For any business that relies on AI, this means:
- Diversify your model providers (don’t bet everything on a single vendor)
- Seriously evaluate open weight models for use cases where control and privacy matter
- Budget for AI as an infrastructure line item, not a free bonus
If You’re in Europe
Digital sovereignty takes on new meaning. When the best American models are proprietary and the best Chinese models are open, Europe is caught in the middle. Mistral remains the only significant European player in the race, with Mistral Small 4 under Apache 2.0 as a credible option. But the investment gap is staggering: Meta is spending over $115 billion on AI capex this year. The EU AI Act’s total budget to support the European ecosystem is measured in hundreds of millions. Two orders of magnitude apart.
The question for European companies is no longer “open source or proprietary?” but “which dependency are we willing to risk?”
Key Takeaways
- Meta launched Muse Spark, its first proprietary AI model, breaking with three years of open source strategy built around Llama
- $14.3 billion invested in Scale AI and Alexandr Wang to steer this new direction through Meta Superintelligence Labs
- The model is competitive but not dominant: strong in reasoning and healthcare, behind in coding and agentic tasks
- Access is the most restrictive on the market: no public API, no open weights, available only within Meta products
- Open source AI isn’t dying — it’s migrating to Chinese labs (GLM-5.1, Qwen) and Google (Gemma 4)
- Meta promises to open-source future versions of Muse, but that promise remains to be verified — especially after the Llama 4 precedent
The message is clear: in 2026, frontier AI is a business, not a public good. Companies and developers who counted on Big Tech’s generosity need to adapt. The good news is that alternatives exist — you just need to know where to look.
FAQ
Is Muse Spark better than ChatGPT or Claude?
Not overall. Muse Spark is competitive but doesn’t surpass Claude Opus 4.6 or GPT-5.4 on most benchmarks. It excels in healthcare (HealthBench Hard: 42.8%) but falls behind in advanced reasoning and coding.
Will Meta continue developing Llama?
Yes, Meta says Llama remains an active project. Muse is a new model family developed in parallel by the Superintelligence Labs. But the massive investment in Muse suggests it’s now the priority.
Why is Meta abandoning open source?
Meta hasn’t officially “abandoned” open source — the company says it hopes to open-source future versions of Muse. But the reality is that the race to superintelligence, the Llama 4 failure, and $115-135 billion in 2026 capex make the open source approach economically difficult for frontier models.
