Meta Put a 30B Model on Your Laptop, and Zuckerberg Used the Launch to Pick a Fight
Muse Glimmer is free to download, runs on a single consumer graphics card, and beats bigger rivals on agent tasks. The essay that came with it is the more interesting part.
Meta released Muse Glimmer on Monday, a 30 billion parameter model you can download for free and run on one ordinary graphics card. It ships under the Apache 2.0 licence, which means you can use it commercially, modify it, and build products on it without asking anyone. For a company that spent the last year keeping its best models to itself, that is a noticeable turn.
The practical numbers matter more than the licence. Compressed to roughly 4 bit, the model needs under 20 GB of memory, which puts it inside reach of a gaming PC or a well specced Mac. It handles text, images and video, holds a 131,000 token context (roughly, how much it can keep in mind at once), and speaks over 100 languages. Hugging Face shipped same day support in transformers, llama.cpp and vLLM, so getting it running is a one line install rather than a weekend project.
On agent benchmarks, which measure whether a model can actually use tools and finish multi step jobs, Glimmer does well for its size: 75.5 on MCP Atlas against 54.2 for Gemma4-31B, and 51.2 on SWE-Bench Pro, a software repair test. It loses to Qwen3.6-27B on several coding and desktop benchmarks, so this is not a clean sweep. Worth knowing before you get excited: on one safety measure, resistance to prompt injection attacks, Glimmer sits in the middle of the pack.
Meta stopped publishing open weights when its Llama line fell behind, and this is the first open model from its new Superintelligence Labs. In an essay published alongside the release, Mark Zuckerberg argues that distilling other labs’ models, that is, training a small model to imitate a bigger one, is legitimate practice rather than theft, and calls for fewer restrictions on American labs. That is a direct swipe at OpenAI and Anthropic, both of whom have argued the opposite. An open weight version of the larger Muse Spark 1.2 is expected to follow, according to the Wall Street Journal.
What this means for you: if you have been curious about running an AI model on your own machine, with nothing leaving your computer, this is one of the easiest starting points yet. Install llama.cpp, pull the GGUF file, and you have a private assistant that costs nothing per question. If you already run local models, the interesting part is that Glimmer was tuned for agent work, so it can hold a tool loop together rather than just chat. For everyone else, the takeaway is simpler: the gap between what you can run at home and what you rent from a lab keeps shrinking.
Sources
OpenAI Built a Model That Writes Exploits, and Handed It to a Short List of Companies
GPT-5.6-Cyber answers 95 percent of advanced hacking requests where the normal model answers 1.5 percent. It already found a real Chrome vulnerability and over 400 kernel bugs.
