Zuckerberg defends distillation, the training technique Trump wants to restrict

This is a developing story. An earlier version was published at 15:43 PDT under the headline 'Meta picks a side in the AI war.'

Mark Zuckerberg published a lengthy essay on Monday arguing advanced AI should be "broadly distributed" rather than concentrated in "a select few companies, institutions or governments" (the essay, via about.fb.com, also hosted at meta.com/thefutureisforeveryone). 404 Media's headline pegs the piece at 6,500 words (404 Media). The same morning, Meta released Muse Glimmer, a lightweight AI model that runs on a personal computer, and gave developers access to a more capable model, Muse Spark 1.2, under what the company describes as a permissive open-source license (ABC News, AP).

Read together, the product release and the CEO essay are one move. Meta is asking the public, the developers, and Washington to evaluate the AI race on the axis where the company has a structural advantage: distribution, code access, and open developer ecosystems. The wire's characterization, in a single phrase, is that Meta "has struggled to keep up" with closed-source rivals OpenAI, Anthropic, and Google (ABC News). That is a reporter's framing, not an analyst consensus, but it is the through-line. A model the public can run on a personal computer will not beat the current frontier systems on most benchmarks, and Meta's open-weight releases are unlikely to change that score. What they can do is change what the question is.

Zuckerberg makes the reframing explicit in the essay. "Most other labs are focused on building AI for companies, governments, or other institutions," he writes. "So if those labs lead, then the balance of power will favor larger institutions over individuals" (ABC News). The line is a veiled reference to OpenAI, Anthropic, and Google, which have moved toward enterprise and government contracts as their commercial anchor. Meta's counter, in distribution terms, is the global base of Facebook, Instagram, and WhatsApp users and the developer community that has built on PyTorch and the company's prior open-weight models.

The policy asks in the essay follow the same logic. Zuckerberg calls for the United States and its allies to expand energy capacity and infrastructure to stay ahead of China on AI buildout, and to reconsider so-called distillation rules (Meta). Distillation is a training technique in which a smaller model learns from a larger one, and it is the central allegation in Anthropic's earlier-in-2026 complaint that China's DeepSeek extracted features from US models. The Trump administration has vowed to crack down on the practice. By defending distillation as legitimate, Meta is arguing for rules that would benefit open-weight developers, including itself, more than the frontier closed labs currently complaining about it.

This is the structural argument the wire tends to flatten. The essay is not a corporate philosophy statement in a vacuum; it is timed to coincide with a model release, previewed by Zuckerberg on X (@finkd), and it lands inside two live fights in Washington: one over energy buildout for AI data centers, and one over what kinds of training cross-border AI companies can do. The 404 Media headline calling the essay "deranged" is one labeled critical perspective, not a neutral description (404 Media). The more durable question is whether the public-policy fight over open-versus-closed AI is, at this point, also a fight over who gets to define the terms.

The reader is not asked to weigh whether open-source AI is virtuous. They are asked whether the axis Meta wants to measure the race on (distribution, developer access, and the policy levers that follow) is the right one, or whether the closed-model labs' axis (benchmarks, capability, and enterprise revenue) will keep winning. Meta's bet, made explicit in this week's release, is that it can change which axis counts.