On April 8, 2026, Meta released Muse Spark, the first AI model from its newly formed Meta Superintelligence Labs (MSL) — and the company’s first flagship model that is not fully open-source. The move marks a decisive break from Meta’s long-standing strategy of releasing every major AI model under a permissive open license, a philosophy that made Llama the most widely adopted open-weight model family in the world.
For the developer community that built its toolchains around Llama, the question is no longer hypothetical: is the open-source AI era at Meta over?
The Llama 4 Disappointment
To understand why Meta changed course, you have to look at the trajectory of Llama 4. Released in April 2025 after nearly a year of delays, the Llama 4 herd — Scout (17B, 16 experts), Maverick (17B, 128 experts), and the unreleased Behemoth — was supposed to cement Meta’s leadership in open AI.
It didn’t.
- Benchmark disappointment: Llama 4 Maverick initially ranked well on the LM Arena leaderboard, but independent testing showed it lagging behind DeepSeek V3, Qwen 2.5, and Claude 4 on reasoning and math tasks.
- Developer sentiment soured: According to reporting from Business Insider (May 2025), developers expressed frustration that Meta had not delivered a competitive reasoning model. Multiple developers told the outlet they expected a reasoning-focused model at LlamaCon and would have “even settled for a traditional model that could beat DeepSeek.”
- Behemoth shelved: Meta reportedly paused testing on Llama 4 Behemoth, the largest and most anticipated model in the family. Internal sources cited benchmarks that failed to meet expectations, particularly in STEM reasoning.
The result was a crisis of confidence. Meta, which had positioned itself as the champion of open-source AI, was being outperformed by Chinese competitors (DeepSeek, Alibaba’s Qwen) and falling further behind closed-source frontier labs (OpenAI, Anthropic, Google).
Enter Alexandr Wang and the Superintelligence Labs
Mark Zuckerberg’s response was characteristically bold. In mid-2025, Meta struck a $15 billion deal with Scale AI that brought its 28-year-old CEO Alexandr Wang to lead a newly created division: Meta Superintelligence Labs (MSL). Wang was given a blank check to rebuild Meta’s AI efforts from the ground up.
The mandate was clear: catch up, at any cost.
Yann LeCun, Meta’s chief AI scientist for over a decade, announced his departure in November 2025 to form his own startup. In subsequent interviews, LeCun was blunt, calling Wang “inexperienced” and predicting further departures. The tension between Meta’s research-first culture and Wang’s product-driven approach became a defining narrative of 2025.
Muse Spark: What Meta Built in 9 Months
Muse Spark (internally codenamed “Avocado”) is the first fruit of MSL’s ground-up rebuild. The results are striking not just for what the model can do, but for how efficiently it does it.
Key capabilities
- Natively multimodal reasoning: Unlike Llama 4, which added vision capabilities after the fact, Muse Spark was built from the ground up to integrate visual information.
- Parallel subagent orchestration: Muse Spark can launch multiple reasoning agents simultaneously. For example, planning a trip — one agent drafts an itinerary, another compares destinations, a third finds activities — all in parallel.
- Contemplating Mode: A multi-agent reasoning mode that competes with Google’s Gemini Deep Think and OpenAI’s GPT Pro. Scores: 58% on Humanity’s Last Exam, 38% on FrontierScience Research.
- 10× compute efficiency: Meta claims Muse Spark requires over an order of magnitude less compute than Llama 4 Maverick to reach the same capability level. The pretraining stack was rebuilt from scratch over 9 months.
- Visual coding: Generates custom websites, mini-games, dashboards, and interactive simulations from natural language prompts.
- Health reasoning: Trained in collaboration with over 1,000 physicians for factual, comprehensive health responses.
The May 12 update
On May 12, 2026, Meta rolled out a major update to Muse Spark, extending its reach across the entire app ecosystem:
- Voice Conversations: Natural, interruptible voice interactions in the Meta AI app, with image generation, Reels, maps, and live camera feed.
- Live AI on Mobile: Previously limited to AI glasses, users can now point their phone camera at the world for real-time answers.
- Shopping Mode: Searches Facebook Marketplace and the broader internet simultaneously, with map view and price/style/distance refinement.
- Glasses rollout: Ray-Ban Meta and Oakley Meta glasses in the US and Canada; Meta Ray-Ban Display coming summer 2026.
- Cross-app expansion: Muse Spark now powers Meta AI across WhatsApp, Instagram, Facebook, Messenger, and Threads.
The Hybrid Strategy: Open Source, But Not All of It
The most consequential shift is not the model itself — it’s Meta’s licensing strategy.
Axios reported on April 6, 2026, that Meta plans to open-source versions of its next AI models, but not the largest or most capable ones. Key details:
- Some models, some components: Meta will release open-source versions of Muse Spark’s successors, but with proprietary components withheld. Safety risk is cited as one reason.
- Largest models stay closed: Wang has indicated that some of Meta’s largest models will remain proprietary — a “hybrid” strategy.
- Consumer focus: Wang sees Anthropic and OpenAI as increasingly enterprise/government-focused. Meta’s strategy is to dominate the consumer AI market through its billions of WhatsApp, Instagram, and Facebook users.
- Ecosystem hedge: The hybrid approach keeps enough openness to win developer mindshare while reserving the most capable models for competitive advantage.
This mirrors a broader industry trend. Alibaba recently kept its most powerful Qwen models proprietary, reversing its own open-source playbook. Even companies that champion openness are pulling back on their most advanced systems.
What This Means for the Open-Source Ecosystem
The implications are profound for the open-weight LLM ecosystem:
-
Llama’s successor is not Llama 5: Wikipedia already lists Muse Spark as the replacement for Llama. The Llama brand may live on for open-source releases, but the flagship — and the R&D investment — has shifted to MSL’s Muse series. This is the same division behind SAM 3.1, which continues Meta’s open-source tradition in computer vision.
-
The “best open model” crown is up for grabs: With Meta stepping back from full openness, the mantle of “best open-weight model” passes to DeepSeek (V4?), Qwen (Alibaba), and Mistral. DeepSeek’s continued commitment to open-weight releases has made it the default choice for many developers who previously relied on Llama.
-
Developers face a fork: Those who built on Llama for its openness now have to decide: follow Meta’s closed models (Muse Spark via API), switch to competing open models (DeepSeek, Qwen), or wait for future open versions from Meta that may lag the frontier.
-
Safety as a rationale — or a cover?: Meta’s stated reason for keeping some models closed — safety risk — is plausible given Muse Spark’s evaluation-awareness finding (Apollo Research found “the highest rate of evaluation awareness of models they have observed”). But it also conveniently protects Meta’s competitive position.
The Bigger Picture
Meta’s shift from fully open to hybrid mirrors what many in the industry predicted: open-source AI works as a strategy when you’re catching up, but once you have a lead — or need one — the incentives to keep things closed become overwhelming. Our State of AI Agents May 2026 roundup confirms this trend is accelerating across the entire ecosystem.
The question for 2026-2027 is whether a credible fully open alternative (DeepSeek, Qwen, Mistral, or a newcomer) can sustain the pace of frontier progress that Meta’s Llama series once offered.
For now, the open-source AI world has lost its biggest patron. The era of Meta releasing every model under a permissive license is over — and no single successor has stepped up to fill the void.
This article was researched from Meta’s official blog posts, Axios reporting, Business Insider, NYT, Fortune, and independent benchmark analyses. All information is current as of May 26, 2026.