In June 2025, Mark Zuckerberg signed a check for $14.3 billion — a 49% stake in Scale AI — and handed the keys to Meta’s AI future to Alexandr Wang, Scale’s 28-year-old co-founder and CEO. Nine months later, Meta shipped Muse Spark, its first flagship AI model that is deliberately, unapologetically proprietary. The Llama era is over.
This is the inside story of how Meta went from open-source AI’s loudest champion to its most surprising defector — and what it means for the millions of developers, startups, and enterprises that built their AI stack on Llama.
The Llama 4 Disaster That Broke Meta’s Open-Source Faith
The seeds of this reversal were sown in April 2025, when Meta launched Llama 4 — Scout, Maverick, and a preview of Behemoth — with fanfare about “natively multimodal” models and Mixture-of-Experts architectures. The reality was brutal.
| Benchmark | Meta Claimed | Independent Score |
|---|---|---|
| Artificial Analysis Intelligence Index v4 | Not disclosed | 18 (vs GPT-5.4 at 57) |
| Community Sentiment | “Next frontier” | “Underwhelming / mid” |
Llama 4 Maverick scored 18 on the Intelligence Index — below models with half its training budget. The community that had championed Llama 2 and Llama 3 turned openly hostile. Allegations of benchmark gaming surfaced: Meta had reportedly applied “post-training tweaks designed to inflate scores.” The frontier conversation moved to GPT-5, Gemini 3, and Claude Opus 4.5. Llama 4 Behemoth was delayed indefinitely and became vaporware.
“Llama 4 Maverick scored 18 on Artificial Analysis’s Intelligence Index, placing it below models half its training budget. The gap between 18 and 52 — Muse Spark’s score — is wider than the gap between Muse Spark and the current #1.” — WhatLLM.org analysis
The $15 Billion Scorched-Earth Rebuild
Zuckerberg didn’t iterate. He burned it down.
| Event | Date |
|---|---|
| Alexandr Wang hired as Meta’s first Chief AI Officer | June 2025 |
| $14.3B for 49% of Scale AI | June 2025 |
| Meta Superintelligence Labs (MSL) formed | Mid-2025 |
| Aggressive poaching from OpenAI, Anthropic, DeepMind ($100M–$300M comp packages) | Mid-2025 |
| Yann LeCun leaves, calls new leadership “inexperienced” | Early 2026 |
| Muse Spark announced — closed-source, first MSL model | April 8, 2026 |
The scope of the rebuild was unprecedented:
- New pretraining architecture — built from scratch, not a Llama derivative
- New data curation pipelines — led by Wang’s data-centric approach from Scale AI
- New RL systems — including “thinking time penalties” and “thought compression”
- New culture — “Demo, don’t memo” (working prototypes over research papers)
The result? Muse Spark scored 52 on the Intelligence Index — the largest single-generation jump ever recorded by a major lab. It now sits top-5, behind only GPT-5.4 (57) and Gemini 3.1 Pro (57), tied with Claude Opus 4.6 (53).
What Muse Spark Actually Does
Muse Spark is a natively multimodal reasoning model with three inference modes:
- Instant — fast answers to simple queries
- Thinking — extended single-chain reasoning
- Contemplating — multi-agent parallel reasoning: launches sub-agents that tackle different parts of a complex request simultaneously, then synthesizes outputs
Key Benchmark Performance
| Benchmark | Muse Spark | GPT-5.4 | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Intelligence Index v4 | 52 | 57 | 53 | 57 |
| HealthBench Hard | 42.8% | 40.1% | 14.8% | — |
| GPQA Diamond | 89.5% | 92.8% | 92.7% | 94.3% |
| CharXiv Reasoning | 86.4% | 82.8% | — | 80.2% |
| ARC AGI 2 | 42.5% | 76.1% | — | 76.5% |
| SWE-Bench Verified | 77.4% | — | — | — |
Muse Spark leads on health reasoning (42.8% on HealthBench Hard, trained with input from 1,000+ physicians) and multimodal chart understanding (CharXiv 86.4%). Its weakness is abstract reasoning (ARC AGI 2: 42.5% vs 76%+ for leaders).
The model also claims order-of-magnitude efficiency — reaching or exceeding Llama 4 Maverick’s capabilities with over 10x less compute — and uses far fewer tokens during evaluation (~58M output tokens vs Claude Opus 4.6’s 157M).
The Closed-Source Pivot That Broke Trust
Here’s the rub: there is no migration path from Llama to Muse Spark.
| Aspect | Llama | Muse Spark |
|---|---|---|
| Licensing | Open-weight download | Fully proprietary |
| Hosting | Self-host or cloud | Cloud-only API |
| Fine-tuning | Full model access | No weights available |
| Pricing | Free | Paid API (private preview) |
| Future | Maintenance mode | Active development |
Meta’s official statement: “Llama line will continue separately” — but the consensus is clear: Llama is in maintenance mode. All significant resources flow to MSL and Muse Spark.
Alexandr Wang promised: “We plan to open-source future versions.” But the community is deeply skeptical — many view this as a deflection, noting Meta conveniently closed the gates once it had something worth protecting.
“Meta’s move away from being the leading U.S. champion of open weights is a significant loss for the developer community.” — Andrew Ng, The Batch
Where Llama Developers Go From Here
The Llama ecosystem reached 1.2 billion downloads before this pivot. Thousands of companies built on it. The migration costs are significant:
- Rewriting vendor-specific APIs for new model providers
- Adapting proprietary training data that was aligned to Llama architectures
- Rebuilding custom tooling for fine-tuning, deployment, and evaluation
Viable Alternatives
- Continue using existing Llama models — available on major cloud providers, but increasingly behind frontier
- Switch to open-source competitors — Mistral, DeepSeek, Alibaba’s Qwen (all actively maintained)
- Llama forks that will continue independently:
- llama.cpp — popular C++ inference engine
- ik_llama.cpp — performance fork
- OpenLLaMA — Apache-licensed reproduction
- Adopt proprietary APIs — from OpenAI, Anthropic, Google, or Meta’s own Muse Spark
The Bigger Picture: Meta’s New AI Strategy
Meta’s pivot is not just about models — it’s about business model transformation.
Muse Spark is rolling out across WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses — a distribution network of 3.2 billion daily users. Unlike OpenAI or Anthropic, Meta is not primarily selling API access to developers. It’s building an AI layer on top of the largest social platform in history.
The real product? Personal superintelligence — an AI that knows your social graph, your shopping preferences, your health concerns, and your physical surroundings (through camera glasses). Privacy advocates are alarmed: TechCrunch noted Meta “doesn’t explicitly say personal info will be used,” but given Meta’s history of training on public user data, the expectation is clear.
Meta’s stock rose 9% on launch day. Investors viewed Muse Spark as proof the $14.3 billion bet produced real results.
What This Means for Open-Source AI
The loss of Meta as an open-weight champion is a structural blow to the open-source AI ecosystem. Meta was unique: a major US tech company with the resources to compete at the frontier, openly sharing weights. Now:
- No major US corporation is releasing frontier open-weight models
- Mistral (France), DeepSeek (China), and Qwen (China) become the primary open-source leaders
- OpenAI’s GPT-OSS release and Google’s Gemma are smaller-scale efforts
- The regulatory conversation around open-source AI safety loses its most visible example
The Verdict
Muse Spark puts Meta back in the AI race technically — but at the cost of the community trust that made Llama a phenomenon. The 18→52 Intelligence Index jump is impressive, but it was achieved by closing the doors that hundreds of thousands of developers walked through.
For the developer community, the message is clear: don’t build your infrastructure on a corporation’s benevolence. The open-source AI ecosystem is only as open as the next earnings call allows it to be.
Coverage aggregated from Meta’s official Muse Spark announcement, The New Stack, Forbes, WhatLLM.org, and CNBC.