Muse Glimmer ships under Apache 2.0. Every Llama before it carried a scale veto Meta could aim at Google. Dropping that clause prices what an American open model is worth in August 2026.
Mark Zuckerberg published 6,500 words on Monday about balance of power and personal superintelligence. The document that moves money is the one-line license notice on Muse Glimmer. Meta spent three years attaching a 700-million-user gate, a naming rule and an acceptable use policy to every open release. The new model carries none of them, because terms are something you can only charge when the buyer has no substitute.
Zuckerberg’s essay ran 14 pages and covered recursive self-improvement, superintelligent lawyers and his eight-year-old daughter’s coding habit. The Wall Street Journal spent its Tuesday newsletter on the semantics, pointing out that open weights and open source are different things and that Meta benefits from blurring them. Both are right and both skip the number.
Meta released the weights for a 30-billion-parameter model under Apache 2.0. Read the four license notices Meta has attached to its open models since 2023 and you get a price series for something nobody quotes: the terms an American lab can impose on a developer who downloads its weights. That price went to zero this week.
What Happened
On Monday, August 10, Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter dense multimodal model built for local agent work, and published the weights on Hugging Face under Apache 2.0. Meta shipped full-precision BF16 weights, two 4-bit quantized variants, a speculative-decoding drafter and the perception encoder, with a 4-bit deployment envelope of roughly 20 gigabytes. The pitch is a Mac or PC with one consumer GPU. Meta did not release the training data or the training code.
Zuckerberg paired the release with “The Future is for Everyone,” a 6,500-word essay arguing that concentrating AI in a few institutions is the actual safety risk. He said Meta will resume releasing some open models, that Meta will open the weights of Muse Spark 1.2, and that the United States needs to reconsider its policies on distillation and training data if American open models are going to lead. His line on distillation: it matters to protect the principle that you can learn from anything you can observe.
He also announced a $1 billion Future Is For Everyone Fund for data center host communities, and a governance change giving Meta’s independent board of directors the power to approve the safety criteria for model releases and to review whether each release meets them.
Muse Spark 1.2 has no publication date and no published license.
The Backstory
Meta shipped Llama 4 Scout and Maverick on April 5, 2025. The launch went badly. Meta submitted a tuned variant to a leaderboard while releasing a weaker one, Yann LeCun later described the benchmark results as fudged, Chris Cox lost oversight of the AI division, and Behemoth never shipped at all.
Zuckerberg then bought a new leadership team. Meta put roughly $14.3 billion into Scale AI and installed its 28-year-old founder Alexandr Wang as chief AI officer on June 30, 2025, with Nat Friedman on AI products. By December, CNBC reported that Meta had pivoted toward a proprietary flagship. Meta cut the FAIR lab in October 2025 and LeCun resigned. Muse Spark arrived on April 8, 2026 as the new lab’s first model, and Meta kept it closed.
That is a 16-month gap in open releases, from Llama 4 to Muse Glimmer. Zuckerberg’s own essay says a two-month lead is worth a lot in this industry.
The open ecosystem did not wait. DeepSeek ships under MIT. Much of Alibaba’s Qwen family ships under Apache 2.0. Moonshot released Kimi K3 in July at 2.8 trillion parameters, which BMA covered as the point where open weights stopped being a discount tier. By June, Chinese-origin models were taking more routed tokens than American ones, a figure we used in our piece on the Washington open-source fight. Solo founders have been swapping Claude for Chinese weights on margin math for most of a year.
Meta walked back into a market where its former product is the third choice and the two ahead of it charge nothing for the paperwork.
The Plan
Look at what Meta has asked developers to accept in exchange for free weights.
Llama 2 in July 2023 came with four Meta-specific obligations: a clause requiring anyone with more than 700 million monthly active users to request a separate license at Meta’s sole discretion, a ban on using Llama outputs to improve any other large language model, an attribution notice, and Meta’s acceptable use policy. Llama 3 in April 2024 kept all four and added a fifth, requiring derivative models to carry “Llama 3” at the front of their name. Llama 4 in April 2025 dropped the distillation ban but held the user gate, the naming rule, the acceptable use policy and a requirement to display “Built with Llama” on your site.
Muse Glimmer carries none of it. Apache 2.0 gives Meta no user gate, no naming rule, no acceptable use policy, and no lever to withdraw the grant.
The 700-million-user clause was never aimed at startups. It was aimed at Google, Apple, Amazon, ByteDance and a handful of others, and it let Meta hand a rival free R&D while holding a switch. Meta has now given that switch up on a model good enough to run an agent loop on a laptop.
Meta did not find a philosophy. Meta found a market price. Qwen and DeepSeek made restrictive terms a thing you can no longer sell, so Meta stopped listing them.
The Business Model Angle
Meta can zero-price a model because Meta has never earned a dollar from one.
In 2025, Meta booked $196.18 billion of advertising revenue against $200.97 billion of total revenue, about 97.6%, a ratio we used to explain why Meta’s AI moderation keeps deleting real businesses. Advertisers pay Meta. Developers do not. Neither do the labs Meta competes with for talent. When a model has no revenue line, giving it away costs Meta nothing on the income statement and costs OpenAI and Anthropic real pricing power on theirs.
That asymmetry runs through every policy ask in the essay. Zuckerberg wants distillation protected, training-data friction reduced, and no restrictions on foreign open models, while keeping silicon export controls in place. A lab that sells tokens cannot make those arguments without arguing against its own book. We took apart how each player’s position on open source decodes as their P&L last week; Monday’s essay is the clearest version of the trade yet, because Meta is the only participant proposing to abolish property rights in models while holding none worth defending.
Two further items in the essay deserve more attention than the manifesto is getting.
The first is a pricing announcement in the middle of a philosophy document. Zuckerberg writes that paid users will buy compute through a dynamic auction that gives everyone the lowest price for the intelligence they use. Every rival sells subscriptions or per-token rates. Meta plans to sell inference the way it sells ad impressions, which is the one mechanism Meta has run at scale for fifteen years and the reason the ad model works at all. Nobody else in this market has that muscle.
The second is where Glimmer runs. A 30-billion model in a 20-gigabyte envelope on a consumer GPU sets a ceiling on what anyone can charge for the tasks it handles. Companies budgeting agent workloads have been watching token costs eat the savings all year, and a competent local model removes the meter from a growing band of them. Google, Microsoft and Amazon all sell cloud inference. OpenAI and Anthropic sell tokens. Meta sells neither, so pushing work onto the user’s own hardware costs Meta nothing and takes revenue from five companies at once.
The Risk
Apache 2.0 cuts both ways, and the second edge is sharp.
Meta cannot revoke this grant. Google can fine-tune Muse Glimmer into Android. ByteDance can ship it. A Chinese lab can distill it, which Zuckerberg has now argued should be legal as a matter of principle. Glimmer is itself distilled from Muse Spark, so Meta is publicly demonstrating the technique it wants Washington to protect on an artifact it paid for. Apply the same principle to Muse Spark 1.2 the day its weights land and the ecosystem gets a frontier-adjacent model on Meta’s compute bill.
That bill is the real exposure. Meta guided 2026 capital expenditures to $130 billion to $145 billion, against $72.2 billion in 2025. Second-quarter capex hit $31.08 billion, revenue grew 28% to $60.8 billion, and free cash flow fell to $784 million from $8.55 billion a year earlier. Meta is financing part of the buildout through structures we examined in the El Paso data center deal with BlackRock. The output of that spend now ships under a license that forbids Meta from charging for it, and the return has to show up somewhere else: ad conversion, agent adoption, device attach, hiring. None of those has a line item that reconciles to $145 billion.
Meta also just handed its board a veto over releases. An independent director approving safety criteria is a supply constraint that did not exist last week, and Zuckerberg wrote the sentence himself.
The honest counterargument is that Meta has done this before and won. React and PyTorch cost Meta nothing in license revenue and bought standards influence, recruiting and an ecosystem shaped around Meta’s hardware assumptions. Restrictive terms would have killed both. The 700-million-user clause was never enforced against anyone as far as the public record shows, so surrendering it may cost Meta nothing it was actually collecting. And OpenAI shipped gpt-oss under Apache 2.0 a year ago while continuing to sell models, so a permissive license is not by itself a confession.
Quick Questions
Is Muse Glimmer open source? No. Meta released the weights and withheld the training data and the training code. Apache 2.0 covers what Meta published, which is a large file of numbers, not a build you can reproduce.
Does the Apache 2.0 license change anything for a small business? For most, no. The 700-million-user clause never touched you. The naming rule and the “Built with Llama” requirement did, and both are gone, so you can ship Glimmer inside your product without branding Meta into your interface.
Why does Zuckerberg want distillation protected? Because Meta buys more from the technique than it loses. Meta distilled Glimmer from its own flagship, and a rule against learning from model outputs would raise Meta’s training costs while protecting the labs that sell access.
What should I watch next? The license line on Muse Spark 1.2. Apache 2.0 on a laptop model is cheap. Apache 2.0 on the model that powers Meta AI would be the actual commitment.
The Business Model Analyst Take
Ignore the manifesto and read the notice file. Companies tell you what an asset is worth by what they charge for it, and Meta has now charged nothing four different ways in three years, ending at nothing at all.
For your own business, the transferable lesson is about terms rather than price. Meta could dictate license conditions in 2023 because there was no substitute at that quality. It cannot in 2026, and the moment the substitute appeared, every clause Meta had written became a reason to pick someone else. If your contract carries conditions your customers accept only because they have nowhere else to go, those conditions are not part of your product. They are a temporary tax, and you will find out the week a competitor stops charging it.
The wider read is that model quality has stopped functioning as an asset. Zuckerberg says a two-month lead is valuable and innovations get absorbed within months, which is another way of saying the thing costs $145 billion to build and holds value for a season. The companies that survive that arithmetic will be the ones selling something the weights cannot commoditize: distribution, workflow, ground truth, or an auction nobody else can run. Meta is betting on the last two. The labs still selling the model itself are the ones who should read Monday’s essay twice.
