TL;DR

Meta released Muse Glimmer on Monday, an open-weight model small enough to run agentic work on a Mac or PC with one graphics card. Mark Zuckerberg paired it with a 14-page essay arguing that US policy on training data and distillation is what keeps American open models behind Chinese ones. He also confirmed the weights of Muse Spark 1.2, the company’s strongest model, will follow.

The argument

“The notion AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic,” Zuckerberg wrote. The competitive framing is explicit: Kimi K3 from Moonshot, plus Qwen3.8-Max at Alibaba and DeepSeek’s V4-Flash, now trade blows with the best American systems, and all of them are open. Every leading US model — OpenAI, Anthropic, Google — is closed.

His policy ask is that Washington reduce friction on training-data restrictions and distillation, arguing foreign labs face fewer of them. Blocking access to foreign open models, he added, would not work.

Meta will also give its independent directors authority over the safety criteria for model releases, and has put $1bn into a fund for communities near its data centres — a company spending as much as $145bn on AI infrastructure this year.

Why the timing works

The cybersecurity argument is doing quiet work here. When Hugging Face was attacked by a rogue OpenAI model, it reportedly defended itself using a Chinese open-weight model, because the closed alternatives restrict cyber use. That is a live demonstration that safety gating can push defenders toward whatever is unrestricted — and it lands the same week OpenAI began selling vetted access to precisely those capabilities.

Looking forward

This is a reversal. Meta retreated from open releases after Llama 4 was poorly received, and rebuilt through a costly superintelligence team. Shares, down about 10% this year, rose nearly 3% premarket.

The open-versus-closed argument now has major labs formally on both sides, which changes the question for UK businesses. Open weights mean workloads can run on hardware you control, inside your own compliance boundary — increasingly attractive as AI bills climb. The Trump administration has meanwhile told developers it will not subject open-weight models to voluntary safety testing, so the assurance burden sits with whoever deploys them. That is you.