“Learn From Anything You Can Observe”: Zuckerberg Defends AI Distillation, Criticises Closed Labs

Set Trending Topics as a preferred source on Google.

It’s a one-two punch of product and manifesto. On Monday, Meta Superintelligence Labs (MSL) introduced Muse Glimmer – 30 billion parameters, weights freely downloadable on Hugging Face under the permissive Apache 2.0 license. In parallel, CEO Mark Zuckerberg argues in a lengthy position piece that superintelligence should be widely distributed rather than held by a handful of institutions. His core claim: there is no such thing as a single benevolent superintelligence – safety only emerges from a balance of power.

The target of his criticism, though never named, are those labs that keep their strongest models to themselves. The most dangerous scenario, Zuckerberg writes, is not releasing capable models but leading labs holding them back – no matter how much that is justified in the name of responsibility and safety. He also takes issue with how rivals frame alignment: most labs treat it as enforcing a centralised set of values, whereas Meta sees it as aligning an agent with the goals of the individual user.

The contradiction: Muse Spark is closed

That argument runs into a situation Meta created itself. With Muse Spark, the company broke with the Llama tradition of open weights in April 2026: the first frontier model from the newly formed Superintelligence Labs shipped without open weights, as a hosted product. July brought Muse Spark 1.1 along with the first paid Meta Model API – according to the provider, starting at 1.25 US dollars per million input tokens and 4.25 dollars per million output tokens. Version 1.2 followed in early August together with the coding agent Muse Code, again without downloadable weights.

All of this came after a difficult year for Meta’s open-weights strategy: Llama 4 was widely seen as a disappointment, compounded by the debate over a modified, unreleased variant of the model submitted to benchmark leaderboards. Chief scientist Yann LeCun, the face of the company’s open research culture, left Meta in November 2025. The restructuring under Alexandr Wang, who joined Meta through the multi-billion-dollar Scale AI transaction, was explicitly product- and revenue-driven.

Against that backdrop, Monday reads as a course correction, or at least a recalibration. In his essay, Zuckerberg announces that Meta will “soon” resume releasing open source models. Muse Glimmer is the first evidence of that. In addition, Zuckerberg and MSL chief Alexandr Wang used posts on X and Threads to signal open weights for a version of Muse Spark 1.2 – no timeline beyond “in the coming weeks” has been given so far.

What Muse Glimmer can do technically

Muse Glimmer is explicitly not a frontier model but is built for local, always-on agents. Meta trained it from Muse Spark using logit distillation and then post-trained it on agentic tasks. Via a perception encoder it also processes images, such as screenshots, charts or documents; it was trained on data from more than 100 languages.

The trick that makes it run on consumer hardware is quantisation: instead of the more than 55 gigabytes the model would require at full precision, it shrinks to under 20 GB – with no meaningful quality loss on agentic tasks, according to Meta. On top of that comes a lightweight “drafter” model based on DFlash for speculative decoding, which by Meta’s own measurements speeds up generation by a factor of 3.1 on an RTX 5090, 1.8 on an M5 Max and 1.5 on an M4 Max.

As comparison points in its own benchmarks, Meta cites Gemma4-31B and Qwen3.6-27B, i.e. its direct size class. The model is available via Hugging Face and is meant to run through Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM and SGLang, among others; hardware partners are AMD, Arm, Dell, Intel and Nvidia.

How the contradiction can be resolved

There are several readings of the gap between rhetoric and product portfolio, and they don’t rule each other out.

First, the commercial one: Meta’s revenues don’t depend on selling models but on reach and advertising. Small, open models lower the cost of the ecosystem and bind developers to it – while the frontier model is monetised through the API. The division of labour – small and open for local agents, large and closed for paying customers – is economically consistent, even if it grates against the rhetoric.

Second, the geopolitical one: in his essay, Zuckerberg explicitly calls for loosening US rules on training data and distillation so that American open models can catch up with strong Chinese open-weights competition. Export controls on chips, by contrast, should stay. He rejects banning foreign open source models. The market share of Chinese open models is therefore a central motive behind the relaunch.

Third, the regulatory one: the text is also a position paper against proposals to roll out frontier models more slowly or under tighter control. Instead, Zuckerberg proposes giving government agencies access to intermediate training checkpoints, and points to a new governance construct in which Meta’s independent board signs off on the safety criteria for model releases.

Zuckerberg’s arguments: open weights as a question of national competitiveness

Zuckerberg devotes the most detailed part of his essay to the question of which nations lead the development and deployment of advanced AI. Whoever leads there, he argues, will shape the future geopolitical balance of power – and to ensure that the leading systems carry your own values, he writes: “The best way to achieve this is to distribute our systems widely.” Meta wants to contribute to that with open source models, among other things.

His chain of reasoning in detail:

On speed: AI is likely “the most competitive industry in history”, with innovations copied and absorbed within months. But because users always want the most advanced model, even a two-month lead is “incredibly valuable” – an indication of how thin the margin of market leadership is. He puts the policy consequence just as sharply: any policy that slows American model releases, “even by a month”, could seriously jeopardise US leadership while foreign models race ahead.

On leading the open source ecosystem: the US and its allies must also lead on open models, which he expects to account for a large share of global AI use. “Our goal should be for American open source models to be the best globally,” Zuckerberg writes. At the moment he sees American labs at a disadvantage because they have to comply with “many additional restrictions on training data” – US policy, he argues, must reduce that friction.

On distillation: the ability of models to learn from other models is, in his view, a load-bearing principle of the open source ecosystem. All AI models derive from human knowledge, he argues; some have tried to frame distillation as harmful. He wants to protect the principle instead: “you can learn from anything you can observe.” Without it, the US could not lead. Notably, Anthropic has accused Chinese companies of copying Claude models via distillation; meanwhile Apple has, legitimately and for billions in payments, distilled LLMs from Google’s Gemini models.

“For the US to lead in open source, we will need to rethink our policies in several areas, including distillation and data use in training. The ability for models to learn from other models is an important principle of how the open source ecosystem works. All AI models are derived from human knowledge. Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe. This is how the world works, and the US will not be able to lead if we restrict ourselves on this front”, Zuckerberg writes.

Against blocking foreign models: restricting access to foreign open source models is not, in his view, an effective solution. Keeping people and companies away from the best open models – wherever they come from – lowers the quality of the AI available to them and centralises AI instead of putting power in users’ hands.

On security: in cybersecurity, widely deployed open source systems have proven more secure, because more people can find vulnerabilities, harden systems and upgrade to current versions. In the long run, his thesis goes, widely distributed models with strong security capabilities lead to more secure systems – not less secure ones.

On chips and infrastructure: export controls on semiconductors have successfully slowed foreign labs and should, according to Zuckerberg, be continued. When it comes to building energy and data centre capacity, however, he sees the US falling behind: countries such as China are bringing more than a gigawatt of nuclear capacity online every other week.

His conclusion: falling behind in AI overall would almost certainly be a bigger and longer-term national security problem than any individual risk arising during development. The current open source ecosystem is strong, he argues – restricting it would be a mistake.

Open questions

It remains open how far “open” reaches. Muse Glimmer is licensed under Apache 2.0 and is therefore considerably more permissive than the old Llama licenses – but like most providers, Meta does not publish training data or training code, which is why calling it open source in the strict sense remains contested. Equally open: which “version” of Muse Spark 1.2 actually gets open weights, when that will happen, and whether the frontier model will follow suit permanently or whether this remains a one-off gesture.