Meta Abandoned Open Source. Your Llama Stack Just Changed.

Meta launched Muse Spark on April 8, its first model under the new Superintelligence Labs division led by former Scale AI CEO Alexandr Wang. Closed weights. No public download. No Llama branding. VentureBeat titled its coverage “Goodbye, Llama?” which about sums it up.

The company that made open source its entire AI identity shipped a proprietary model and didn’t bother putting the Llama name on it. Bloomberg reported the launch signals a “strategic repositioning” of Meta’s AI efforts from community-driven open weights to commercially controlled products.

If your AI stack, cost projections, or vendor risk assessment assumed Meta would keep releasing frontier-class models for free, that assumption took a hit today.

The Shift at a Glance

	Before Muse Spark	After Muse Spark
Meta’s AI brand	Llama (open-weight, community-driven)	Muse (proprietary, commercially controlled)
Model access	Download weights, run anywhere	API access, Meta controls distribution
License	Llama Community License (free, some restrictions)	Closed/proprietary
Cost to deploy	Hardware only (free weights)	Per-token API pricing (TBD)
Vendor lock-in	None	Yes
Leadership	FAIR team, Yann LeCun era	Superintelligence Labs, Alexandr Wang
Strategic goal	Ecosystem building, developer adoption	Revenue generation, competitive positioning

If three or more items in the right column affect your current deployments, keep reading.

What Muse Spark Actually Is

Muse Spark is a natively multimodal reasoning model with tool use, visual chain-of-thought reasoning, and multi-agent orchestration built in. According to Meta’s announcement, it scores 58% on Humanity’s Last Exam and 38% on FrontierScience Research in its contemplating mode. Those numbers put it in competitive range with GPT-5.3, Claude Opus 4, and Gemini 3.1 Ultra. It doesn’t surpass any of them across the board.

The efficiency story is more notable. Meta claims Muse Spark uses “over an order of magnitude less compute” than Llama 4 Maverick for equivalent capabilities. That’s a research achievement worth watching. But you can’t watch it up close, because the weights aren’t public.

Distribution is rolling out through the Meta AI app, meta.ai, WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses. Conspicuously absent from that list: any third-party deployment option.

Why This Happened

Three forces pushed Meta from open to closed.

Llama was falling behind. Zuckerberg reportedly grew frustrated that Llama models consistently placed below ChatGPT and Claude on quality benchmarks, despite Meta spending billions on AI compute. Open-source development moves at community pace. Proprietary development moves at the pace of the executive who signs the GPU purchase orders.

Yann LeCun left. Meta’s chief AI scientist departed in late 2025 to launch his own startup focused on world models, per TechCrunch. LeCun had long been critical of the industry’s fixation on LLMs, and the organizational overhaul that followed — including cuts to his FAIR research division and the hiring of Alexandr Wang as chief AI officer — accelerated his exit. With the institutional champion for open AI gone, Meta’s philosophical anchor for open-source commitment left with him.

The money math changed. Meta’s 2026 AI infrastructure budget is $115–135 billion. I covered the scale of that investment when they cut 20% of their workforce to fund it. Giving away the output of that spending when competitors charge per token is a strategy that only works if it generates enough ecosystem value to justify the cost. Meta apparently decided it doesn’t.

The “Hybrid Strategy” Promise

Here’s where it gets murky. Axios reported on April 6 that Meta intends to “partially open-source future models,” with closed versions shipping first for “safety and competitive reasons” and open versions following later.

That framing is doing a lot of work. “Partially open-source” and “later” aren’t commitments you can build infrastructure plans around.

At LlamaCon in April 2025, Meta celebrated surpassing 1 billion Llama downloads and announced expanded open-source tooling. Less than a year later, the first model from their flagship new AI division is fully proprietary with no announced timeline for an open-weight release.

My take: the hybrid strategy is probably real. Llama models will probably continue to exist. They probably won’t be the best models Meta produces. And the gap between proprietary Muse and open Llama will determine whether Meta’s open-source AI remains a viable primary strategy or becomes a second-tier option you settle for.

That’s a planning risk you need to price in now.

Three Vendors, Three Diverging Bets

Muse Spark didn’t launch in a vacuum. In the past week alone:

• Google released Gemma 4 under Apache 2.0 — its most permissive open-source license yet. Full commercial rights. 31B Dense model ranking #3 on Chatbot Arena.
• Microsoft shipped MAI models through Foundry, building proprietary alternatives to the OpenAI models it resells.
• Meta launched Muse Spark with closed weights and no third-party deployment path.

Three major AI players. Three completely different strategies for how models reach the market.

Vendor	Strategy	License	Self-Hosting	Lock-In Risk
Google (Gemma 4)	Open-source ecosystem	Apache 2.0	Full support	Low
Microsoft (MAI/Foundry)	Proprietary + reseller	Commercial	Azure only	High
Meta (Muse Spark)	Proprietary, was open	Closed	Not available	High
Meta (Llama, existing)	Open-weight (legacy)	Llama License	Full support	Low (for now)

If you built a model-agnostic stack with routing flexibility, you can shift between these options as strategies evolve. If you hard-coded Llama dependencies expecting Meta to always be the free option, today is your migration planning meeting.

What Should You Do If Your Stack Depends on Llama?

1. Audit your Llama exposure. Identify every workflow, application, and automation running on a Llama model. Categorize by criticality: which ones would break your business if Llama stopped improving?
2. Test Gemma 4 as a drop-in alternative. Google’s Apache 2.0 license is unambiguous. For most workloads running on Llama 3.x or 4 Maverick, Gemma 4 31B Dense is a comparable replacement with a cleaner license and strong benchmark performance.
3. Build a routing abstraction layer. If your code directly calls Llama endpoints with model-specific formatting, add a routing layer that lets you swap models through configuration. This is the single most valuable architectural decision you can make right now.
4. Set a decision deadline, not a migration deadline. You don’t need to rip out Llama today. Existing Llama models still work. Llama 4 Maverick is genuinely capable. But set a date (90 days is reasonable) to reassess based on what Meta actually ships as open versus proprietary.
5. Watch the gap. If Meta’s next two model releases are both Muse-branded and proprietary, with Llama updates lagging behind, the trend is clear. Don’t wait for an official deprecation announcement. Companies rarely make that announcement. They just stop shipping.

What This Means for Open-Source AI

The broader signal from April 2026: open-source AI is alive, but its champion roster just changed.

Google is now the most aggressive open-source model provider among the hyperscalers. A sentence I wouldn’t have predicted writing a year ago. Gemma 4 under Apache 2.0 is a genuine, unambiguous commitment to open development. Meta held that title for two years. It’s up for grabs now.

The open-weight community built real infrastructure around Llama. Fine-tuning pipelines. Deployment tooling. Compliance frameworks. Entire businesses. That infrastructure doesn’t evaporate because Meta’s strategy shifted, but it does need a new anchor model for the next generation of development.

For enterprises and SMBs running self-hosted AI, the practical advice hasn’t changed: maintain model flexibility, don’t bet your architecture on a single vendor’s licensing philosophy, and keep your abstraction layers current. The geopolitical and vendor risk dimensions I wrote about with the Frontier Model Forum announcement last week compound this. Concentration risk isn’t just about uptime. It’s about strategy shifts you can’t control.

The open-source AI thesis was never “Meta will always give us free models.” It was “capable models will be available without vendor lock-in.” That thesis still holds. The supply chain just shifted, and you need to shift with it.

Your Next Steps

The five-step playbook above covers your technical moves. Start with the Llama audit and Gemma 4 benchmarks this week. Schedule the architecture review and routing layer work within 30 days.

One item for the leadership agenda: three major vendors now run three different model distribution philosophies. That’s a market structure change affecting procurement, architecture, and budget planning. Make sure your decision-makers see it.

Meta built the most valuable open-source AI ecosystem in the industry. Today they signaled that proprietary control matters more. Your job is making sure that decision doesn’t become your problem.

---

Related Reading:

• Gemma 4 Just Made Your API Bill Optional
• Microsoft Is Building AI Without OpenAI
• Your AI Stack Has an Expiration Date
• Frontier Model Forum: AI Geopolitical Risk Just Got Real