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

Meta kicked off Monday with a tidy contradiction. Its Superintelligence Labs released Muse Glimmer, a 30 billion parameter model whose weights are freely downloadable on Hugging Face under Apache 2.0. At the same time, Mark Zuckerberg published a long argument for why superintelligence should be spread out, not kept inside a few companies.

His message is blunt. He says there is no kindly all-powerful AI waiting to keep everyone safe. In his view, the real danger is the opposite: leading labs holding their strongest systems back and calling it responsibility. He also draws a line between Meta’s idea of alignment and the one used by rivals. For Meta, the goal is to line up an agent with the user’s goals, not with some centralised moral code.

The trouble is that Meta has spent the past year making its own open-weights story harder to believe. Muse Spark, the first frontier model from the new Superintelligence Labs, arrived in April 2026 as a hosted product with no downloadable weights. Muse Spark 1.1 followed in July, along with the first paid Meta Model API, priced from $1.25 per million input tokens and $4.25 per million output tokens. Muse Spark 1.2 and the coding agent Muse Code came in early August, again closed.

That makes Monday read less like pure principle and more like a reset. Zuckerberg said Meta will “soon” resume releasing open source models, and Muse Glimmer is the first sign of that. Zuckerberg and Alexandr Wang also used X and Threads to signal open weights for a version of Muse Spark 1.2, though the only timing given so far is “in the coming weeks.”

Glimmer itself is meant for local, always-on agents rather than the frontier. Meta says it was trained from Muse Spark using logit distillation, then post-trained on agentic tasks. It can handle images through a perception encoder, including screenshots, charts and documents, and it was trained on data from more than 100 languages. Meta also says quantisation cuts the model from more than 55 gigabytes at full precision to under 20 GB without meaningful quality loss on agentic tasks, and that a lightweight “drafter” model speeds generation by 3.1x on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max.

The policy pitch is doing a lot of work here. Zuckerberg wants looser US rules on training data and distillation so American open models can keep up with Chinese open-weights systems, while export controls on chips stay in place. He also wants government agencies to see intermediate training checkpoints, and says Meta’s independent board should sign off on safety criteria for releases. In other words: open when it helps Meta, closed when it pays, and regulated when it flatters the argument.