Meta doesn’t sell its most powerful AI models. It gives them away. Llama, and now the Muse-class releases, ship as free downloads under Apache 2.0 while Mark Zuckerberg publishes a 6,500-word manifesto arguing that closed AI is a threat to everyone. That’s a strange position for a profit-driven company to take, and it’s why the story needs unpacking as part of Meta’s wider AI offensive. By the end you’ll be able to separate Meta’s stated philosophy from its commercial incentives, and judge any “open” model on licence, provenance, control and risk rather than narrative.
What does Meta’s 6,500-word manifesto actually argue about closed AI as a structural threat?
The manifesto, “The Future is for Everyone,” is Zuckerberg’s argument that powerful AI “should not be controlled by a handful of companies”, a pointed jab at OpenAI, Anthropic and Google. It rests on three pillars: individual empowerment as the source of prosperity, invention as AI’s purpose, and a balance of power favouring people as the foundation of safety. The stated goal is personal superintelligence for every person. Read it as positioning.
The hard part is that Meta’s own record sits awkwardly against it. The Cambridge Analytica era, and a 2026 employee surveillance program paused after 1,600 staff petitioned, complicate the openness-equals-safety line. And “open” here means open weights, a narrower claim than it sounds. That “open” gap helps explain what came next.
Why did Meta go closed-source with Muse Spark after years of open-source Llama?
In April 2026 Meta Superintelligence Labs did something that looked like a rejection of the manifesto: it shipped Muse Spark, its first closed-source model, with no public weights and API access only. The trigger was adversarial distillation, using a frontier model’s outputs to train cheaper clones. Open frontier weights make that easy. Qwen and DeepSeek had already used open weights to close the capability gap. Meta’s resolution is tiered openness: keep the frontier closed, reopen the lower tiers. Muse Glimmer, a distilled Apache 2.0 model trained on Muse Spark’s own outputs, was the resume-open-source signal.
What’s the difference between open-weight and fully open-source AI models?
“Open source” is doing heavy lifting in Meta’s pitch. What Meta actually releases is open-weight models: the trained parameters you can download, inspect, fine-tune and self-host. Fully open source would also include training data, training code and a reproducible build, which Meta has not released. Apache 2.0 is a clean, permissive licence, friendlier than Llama’s old community licence, but it still doesn’t hand you the recipe. For your business, that’s a cap on auditability and provenance: you can run and tune the model, but you can’t fully verify how it was made. That’s the open-weights distinction in a nutshell.
Why does Meta give away its most powerful AI models for free?
Because foundation models are commoditising anyway, and Meta’s money is not in models. Giving away weights captures developer mindshare and the desktop and tooling layer, then drains the per-token pricing moat that OpenAI and Anthropic live on. The payoff sits in the layers around the model: attention and ads today, and Meta Compute tomorrow. Muse Glimmer and Muse Code are proof points. That is the strategy covered in our piece on selling compute.
What does “commoditise your complement” mean in Meta’s open-source strategy?
That play has a name. Joel Spolsky called it commoditising your complement: lower the price of the thing your customers also need, and demand for your own core product rises. Google did it with Android, commoditising smartphones to sell search and ads. Microsoft did it with Windows. For Meta, the model is the complement; free weights turn it into a cheap commodity, and the more developers build on Meta’s stack, the more valuable its compute and advertising become. Self-hosting swaps per-token pricing for fixed infrastructure cost, which is the fight Meta wants to have.
Meta’s open approach vs OpenAI and Anthropic’s closed approach: which is better for enterprise?
For your business, open versus closed is a trade-off. Meta’s open weights give you control, customisation, data privacy and portability, but you own the operations, security and risk. OpenAI and Anthropic’s closed APIs give you managed support, predictability and vendor-held liability, at the cost of lock-in and opacity. Cost follows the same fork: open trades per-token fees for infrastructure you run, closed bundles everything into the meter. Enterprises have mostly voted with their wallets for closed: by mid-2025, Anthropic led usage at 32%, with Meta’s Llama at 9%. There’s a quiet irony too: Anthropic criticises openness while reportedly leasing around $10 billion of compute from Meta — one thread in Meta’s model and cloud moves. Treat it as a governance decision: closed is managed but locked in; open is portable but you run it, and the trust deficit can cap adoption.
How do Meta’s open models compare with Chinese open-weight models (DeepSeek, Qwen) for enterprise adoption?
If you’ve picked open, the next question is which open. Qwen and DeepSeek often lead on capability and adoption: Qwen overtook Llama in cumulative Hugging Face downloads in late 2025 and nearly doubled it by early 2026, while Chinese models now take a plurality of Hugging Face downloads. Capability isn’t the whole scorecard though. Chinese open-weight models carry data-sovereignty, provenance and geopolitical supply-chain concerns, and because training data and code aren’t shared, you can’t fully verify them. If data sovereignty and Western provenance matter to your compliance regime, Meta’s Apache 2.0 licence sits in a different trust tier, even where capability is comparable.
The giveaway resolves into a single logic: models are the complement, and compute and attention are the product. Meta wins when the model layer is cheap and ubiquitous, so the manifesto is best read as strategy wearing a philosophical frame, one piece of the complete offensive. The choices that follow, open versus closed and Western versus Chinese, are fit-to-risk decisions. Your job is simpler than the debate: audit the licence, the provenance, the control you actually get, and what you own in return.
Frequently Asked Questions
Can my business actually use Meta’s Llama and Muse Glimmer models commercially for free?
Yes. Meta’s Llama and Muse Glimmer releases are open-weight under the Apache 2.0 licence, so a business can download, fine-tune, self-host and deploy them commercially without a per-token fee. You pay only for the infrastructure you run them on. The catch is that “free” refers to the weights, not the training data, so you still carry the operational and compliance burden yourself.
Does using Meta’s open models mean my data gets sent back to Meta?
Not by default. When you self-host an open-weight model, inference runs on your own infrastructure and your prompts and data stay inside your environment. Meta does not receive them simply because you are using the weights. The risk shifts to how you deploy: a managed third-party host or a careless integration could still expose data, so governance and hosting choices matter more than the licence itself.
What are the real hidden costs of running an open model instead of a closed API?
The model is free, but self-hosting is not. You pay for GPUs or cloud compute, MLOps tooling, security hardening, patching, model updates and the people to run it all. A closed API bundles those costs into per-token pricing. The open route usually wins on marginal cost at scale and data control, but it only beats closed pricing if you actually have the engineering capacity to operate it.
Are open-weight models more dangerous than closed models?
Not inherently, but the risk profile differs. Open weights can be stripped of safety guardrails and misused, which is exactly the concern behind Meta’s closed Muse Spark pivot. Closed models centralise that risk behind a vendor’s controls. For an enterprise, the practical question is less about the model’s philosophy and more about what safeguards, monitoring and deployment controls you put around whatever you run.
Is there any catch to Meta’s Apache 2.0 licence?
The main catch is what is not included. Apache 2.0 grants broad commercial use, modification and redistribution, far cleaner than Llama’s earlier community licence and its scale thresholds. But it does not grant training data, training code or reproducible builds, so you cannot fully audit how the model was made. For most enterprises the licence terms are the easy part; the governance gap is the real constraint.
What is adversarial distillation, and why should my business care?
Adversarial distillation is when a rival trains its own model on the outputs of your open model to clone frontier capability cheaply, without the same research spend. It matters to enterprises because it is the reason Meta closed Muse Spark and a reminder that open weights are a strategic asset. For a buyer, it means the free model you adopt can quickly be copied, which reshapes its long-term competitive value.
How far behind closed frontier models are Meta’s open releases really?
The gap has narrowed to the point where it is often measured in months, not years. Muse Glimmer and similar open-weight models now sit close enough to frontier capability for many enterprise tasks, and they can be fine-tuned to beat general-purpose closed models on a specific job. Where closed models still lead is on the hardest reasoning and agentic benchmarks, and on managed, zero-ops convenience.
Will Meta keep giving its models away, or is this a temporary strategy?
The giveaway looks durable because it serves Meta’s commercial interest. As long as commoditising the model layer weakens closed rivals and funnels demand toward Meta Compute and its advertising business, open releases remain strategically rational. The Muse Spark episode shows the line can move at the frontier, so expect the lower tiers to stay open while the most advanced models stay conditional.
How do I actually get started with self-hosting one of Meta’s models?
Start by pulling the weights from Hugging Face or Meta’s release channels, then run them through a local inference stack such as vLLM, Ollama or a managed platform. Match the model size to your hardware and begin with a quantised build before scaling. Treat it like any production system: set up access controls, monitoring, red-teaming and a rollback path before you put real data or customers near it.
Should I worry about Meta’s Cambridge Analytica history when adopting its AI models?
It is a reasonable thing to weigh, but it is not a licence problem. The Cambridge Analytica era raised real questions about Meta’s data governance, and the 2026 employee surveillance lawsuit keeps that scrutiny alive. For an enterprise, the practical response is to treat Meta’s models like any vendor: verify what data leaves your environment, read the licence, and apply your own governance rather than trusting the openness-equals-safety framing.
What does “personal superintelligence” mean in Meta’s manifesto?
It is Meta’s stated end-goal: instead of one central superintelligence controlled by a few institutions, every person gets access to their own capable AI that works on their behalf. It is the philosophical wrapper for the balance-of-power argument. Read it as a positioning statement rather than a product specification, because the manifesto is more specific about the threat than about how the goal is delivered.
