On April 8, 2026, Meta dropped a bomb.
Not the “we have a new model come try it” kind. The kind that made the entire open-source AI community go quiet.
Muse Spark has arrived.
This is Meta’s first brand-new AI model in a year, and the first product from Meta’s newly formed Superintelligence Labs. Leading the charge: Alexandr Wang, poached from Scale AI. Zuckerberg spent $14.3 billion and gave Wang nine months to tear down and rebuild the entire AI infrastructure from scratch.
The result is a powerful model. Also a completely closed-source one.
From Llama to Muse Spark: A Remarkable Pivot
Looking back, Llama’s cumulative downloads crossed 1 billion by early 2026, with about 1 million downloads per day. The developer community built an entire ecosystem on Llama.
Then Meta said: the next one won’t be open.
Muse Spark is a native multimodal reasoning model — with built-in tool use, visual chain-of-thought, and multi-agent orchestration. It now powers Meta AI across all Meta apps, reaching over 3 billion users.
The efficiency gains are real too. Meta claims the new model’s inference cost is a fraction of its predecessor — while running a frontier-tier model. For a company handling billions of interactions daily, this translates to billions of dollars in cost difference.
Benchmarks: Honestly, Not Top-Tier
Meta isn’t claiming to be “the world’s most powerful model” — a notable departure from the Llama 4 overclaiming era.
On the Artificial Intelligence Index v4.0, Muse Spark scores 52, ranking fourth behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6.
But in one area it dominates: medicine.
On HealthBench Hard (open-ended medical queries), Muse Spark scores 42.8 — far ahead of Gemini 3.1 Pro’s 20.6, GPT-5.4’s 40.1, and Grok 4.2’s 20.3. Meta says it collaborated with over 1,000 physicians to organize training data. Healthcare is a clear strategic direction.
Three interaction modes are designed:
- Instant: Quick answers to everyday questions
- Thinking: Multi-step reasoning tasks
- Contemplating: Multiple agents reasoning in parallel, competing — targeting Gemini Deep Think and GPT Pro’s high-end reasoning modes
The Cost of Closed-Source
This is the core of the issue.
Muse Spark is not open-source. No downloading, no self-hosting, no running on your own hardware. It is only available via API to a limited set of partners in private preview. This makes Muse Spark even more closed than OpenAI and Anthropic’s paid models — since at least you can call those via API.
Wang’s own words on the matter:
“Nine months ago, we rebuilt the entire AI stack from scratch. New infrastructure, new architecture, new data pipelines. This is step one. Larger models are in development, with open-source plans for the future.”
“Plans for the future.” Those words sound familiar.
The developer community reaction is split. Some see this as a necessary pivot after Llama 4’s market underperformance — Meta needs to protect its moat with the best offerings. Others feel it’s a case of building a bridge and then closing it: the community helped build the Llama ecosystem, and now that Meta has something good, the door is shut.
User Base Is the Real Weapon
But Meta doesn’t care about developer community sentiment.
Muse Spark is about to be built into Facebook, Instagram, WhatsApp, Messenger, and Meta’s Ray-Ban AI glasses. Think about it: OpenAI and Anthropic are still selling to developers and enterprises one at a time. Meta is stuffing the model into apps that 3 billion people open every day.
That distribution power is more persuasive than any benchmark.
On announcement day, Meta’s stock surged over 9%. Wall Street clearly thinks the $14.3B and nine-month rebuild is worth it.
What About Privacy?
There’s another issue worth watching: Muse Spark requires a Meta account to use. While Meta hasn’t explicitly said it will use personal account data to train AI, the company has a history of using public user data, and it is positioning Muse Spark as a “personal superintelligence” product. When your AI assistant knows all your search history, social interactions, and daily habits, that line gets increasingly blurry.
So What?
Muse Spark isn’t a perfect model. It ranks fourth on general benchmarks, and closed-source means the developer ecosystem built around Llama is left out in the cold. But Meta was never a benchmarks company — it’s a users company.
When 3 billion people chat with Muse Spark on WhatsApp, edit photos on Instagram, and have it write Facebook posts, whether the model is “open source” doesn’t matter to most people.
And for developers, Wang’s promised “future open-source version” will be a question asked every quarter.
Further Reading:
- AI News original: Meta has a competitive AI model but loses its open-source identity
- TechCrunch: Meta debuts the Muse Spark model in a ‘ground-up overhaul’ of its AI
- Meta AI App hits App Store #5: Meta AI app climbs to No. 5
