マーク・ザッカーバーグはAIが「誰のためのもの」と本当に信じているのか？Does Mark Zuckerberg really believe AI is ‘for everyone’?

Meta this week released Glimmer, an open-weight artificial intelligence model that anyone can download and run on their own hardware, reigniting a long-running debate about what "open" actually means in AI and whether the company's rhetoric matches its practice. The release matters because Meta has cast itself as the standard-bearer for accessible AI, and the line it draws between what it gives away and what it keeps locked down shapes the options available to developers, researchers, and smaller firms that cannot build frontier systems on their own.

Glimmer arrived alongside a letter from chief executive Mark Zuckerberg arguing that AI should be broadly available rather than concentrated in the hands of a few companies. The model can be self-hosted, meaning users are not dependent on Meta's servers, subscription tiers, or usage caps to run it. That distinction is meaningful. Open-weight models can be inspected, fine-tuned, and deployed offline, which appeals to organizations with strict privacy requirements, tight cost constraints, or a desire to avoid being tied to a single vendor's pricing and availability.

The contrast the release invites is with Muse Spark, Meta's more powerful model, which stays locked behind the company's own APIs. In practice, that means the most capable version of Meta's technology remains under corporate control, metered and gated, while the freely downloadable option is the weaker one. Observers appear to see a tension in that arrangement: if the strongest tools are reserved for a proprietary channel, then a promise of AI "for everyone" applies mainly to the second tier of capability. Zuckerberg's letter is likely intended to reaffirm Meta's openness credentials, but the simultaneous decision to wall off Muse Spark is what has drawn scrutiny.

It helps to understand the terminology, because the industry uses "open" loosely. An open-weight model means the trained parameters are published, so the system can be downloaded and run independently. That is not the same as fully open-source, which would typically include training data, code, and documentation sufficient to reproduce the model from scratch. Many widely distributed models, including Meta's earlier Llama family, are open-weight but carry licenses that restrict certain commercial uses. Glimmer appears to follow that broad pattern, giving users practical control over deployment without necessarily disclosing everything about how the model was built.

The move also fits a wider strategic split across the sector. Companies such as OpenAI and Anthropic have largely kept their most capable models behind APIs, arguing that controlled access supports safety and sustains revenue. Others, including Mistral and various research groups, have leaned into releasing weights to build developer communities and momentum. Meta has straddled both approaches, using open releases to seed adoption while retaining more advanced or productized systems for its own platforms and paid channels. Offering a downloadable model and a stronger API-only model at the same time is consistent with that dual posture rather than a departure from it.

There are plausible reasons a company might hold back its most powerful model. Frontier systems are expensive to train and serve, and keeping them behind an API allows for usage monitoring, safety guardrails, and monetization that recoups those costs. Publishing the weights of a top-tier model is also difficult to reverse once done. Whether those justifications are persuasive depends on how one reads Meta's messaging: the gap between an expansive "AI for everyone" framing and a tiered release strategy is precisely what critics are pointing to.

For developers and businesses weighing their options, the practical takeaway is that Glimmer expands the pool of self-hostable models, which is genuinely useful for prototyping, on-premises deployment, and cost-sensitive projects. But those needing the highest performance will still route through Meta's paid interface, with the dependencies that implies. The episode is less a single-product story than a reminder that openness in AI exists on a spectrum, and that marketing language and licensing terms do not always align. As more companies release partial or tiered offerings, the meaning of "open" will likely remain contested, and statements about democratizing the technology will continue to be measured against which models are actually handed over and which are kept in reserve.