Zuckerberg reverses course, recommits Meta to open weight AI models
In a 6,500-word essay published August 10, Zuckerberg reversed Meta's April pivot to closed models, promising to 'resume releasing some open source models soon.'
Meta didn't just write an essay. It shipped a 30B open model and opened weights for another.
On August 10, 2026, Mark Zuckerberg published a roughly 6,500-word essay titled 'The Future is for Everyone,' arguing that AI superintelligence should be spread across many hands rather than concentrated inside a handful of labs. In it, he commits Meta to 'resume releasing some open source models soon,' a direct reversal of the company's April pivot to the closed Muse Spark model and of his own July 2025 statement that Meta 'likely won't open source all of its superintelligence models.' The essay names OpenAI, Anthropic, and Google as the labs keeping their best models closed, and holds up China's DeepSeek and Moonshot as the open-weight competition the US is ceding ground to. Meta backed the words with two shipped artifacts the same week: a new 30-billion-parameter open-weight agentic model called Muse Glimmer, and reopened weights for Muse Spark 1.2.
the reversal, dated
Meta's position on open weights has moved twice in thirteen months. In July 2025, Zuckerberg told reporters Meta 'likely won't open source all of its superintelligence models,' a break from the fully open Llama releases that built Meta's AI reputation in the first place. That caution hardened into policy in April 2026, when Meta shipped Muse Spark as a closed, API-only model, its clearest signal yet that the open era was over. The August 10 essay reverses both moves in a single document.
- July 2025: Zuckerberg tells reporters Meta 'likely won't open source all of its superintelligence models.'
- April 2026: Meta ships Muse Spark as a closed, proprietary model, its first major frontier release with no open weights.
- August 10, 2026: Zuckerberg publishes 'The Future is for Everyone' and commits Meta to 'resume releasing some open source models soon.'
what Meta actually shipped
Essays are free. Meta paired this one with two real releases from Meta Superintelligence Labs, the group Zuckerberg built around former Scale AI CEO Alexander Wang. The lab shipped Muse Glimmer, a new 30-billion-parameter open-weight model built for agentic work, and opened the weights on Muse Spark 1.2, walking back the exact model Meta had made proprietary four months earlier.
Meta also announced a governance change alongside the essay: its independent board now holds authority over the safety-release criteria that decide which future models ship open and which stay closed. That moves a call that used to sit with Zuckerberg and his research leadership to a board-level check. Meta has not published the criteria itself, so there is no way yet to see what would trip the switch back to closed.
Read the caveat plainly: this is Zuckerberg's own account of Meta's own motives, published on Meta's own site and pushed through a coordinated round of press interviews the same day. He frames concentrated AI power as the real safety risk and says restricting foreign open models 'is not an effective solution,' positioning open US models as the answer to DeepSeek and Moonshot. That framing arrives right after Meta's closed strategy failed to produce a model that beat those same Chinese labs, or Google, on outside benchmarks. Nothing in the essay is independently audited, and 'resume releasing some open source models soon' commits Meta to no date and no named flagship model, only Glimmer and a version bump on Spark.
Why a build studio cares
This is not abstract industry politics for us. Whether OpenAI, Anthropic, Google, and Meta release open weights or keep everything behind an API decides what we can actually self-host, fine-tune on a client's data, or ship inside an offline agent product. A closed-only market means every AI feature we build carries a permanent per-token bill and a dependency on someone else's uptime. One more open Meta model, even a 30B one built for agent work, is one more option we can run on infrastructure we control.
Next step: read Fortune's coverage of the essay and Meta's interview push, then skim the 630-point Hacker News discussion for the skepticism the essay does not address. If you are weighing an API-only model against something you can actually self-host for a client build, write to us at hello@gattyworks.com.