Enterprise

Meta open-sources Muse Glimmer, a 30B agentic model that runs on a single consumer GPU

Meta Superintelligence Labs shipped an Apache 2.0 model tuned for local, always-on agents — and Zuckerberg used the launch to press Washington on open-source AI policy.

By Asher Glick · Enterprise software desk · 2026-08-10 · 4 min read

Photo: Unsplash / Christian Wiediger — Close-up of a consumer GPU graphics card installed inside a PC

Meta Superintelligence Labs on August 10 released Muse Glimmer, a 30-billion-parameter agentic model, under an Apache 2.0 license on Hugging Face. It’s the most permissive open-weight drop Meta has shipped in over a year, and it’s engineered to fit on a single consumer GPU.

The compression story is the point. Full precision, the model is more than 55 gigabytes; quantized, it’s roughly 17, comfortably inside the 24 or 32 gigabytes of VRAM available on prosumer cards. Glimmer is a distilled variant of Muse Spark 1.2, Meta’s flagship, and it’s tuned for the loops that agent workloads actually run: tool use, multi-step reasoning, failure recovery. It supports text and images and was trained across more than 100 languages. It integrates with OpenClaw, and ships day-one support for llama.cpp and Ollama.

The strategic framing arrived in parallel. Zuckerberg published a 6,500-word, 14-page essay arguing that the U.S. will need to “rethink” its policies in several areas, including distillation and data use in training, if it’s to lead in open source AI. Reuters reported that Meta intends to make the weights of Muse Spark 1.2 publicly available as well, positioning Glimmer as a precursor rather than the ceiling.

Read together, the launch and the essay are one document with two audiences. To developers, it’s a usable local model with a real license. To Washington, it’s a lobbying brief with a working demo attached.

There’s also an investor audience. Meta’s capital expenditure forecast runs up to $145 billion this year, and Superintelligence Labs, which Meta stood up last year, has been under quiet pressure to show output. Chinese open-weight releases from Alibaba, DeepSeek, and Moonshot have competed credibly with U.S. frontier work, and Meta’s return to permissive licensing lands squarely in that competitive frame. Meta stock was up nearly 3% in premarket trading Monday.

The interesting move is that Meta is treating open source as policy leverage, not just distribution strategy. The 2023 Llama 2 release opened a market. This one is trying to shape a rulebook.

Sources