What's this about?
Mark Zuckerberg announced Meta is releasing the *weights* of Muse Glimmer — a 30-billion-parameter AI model tuned for "agentic" work (using tools, taking multi-step actions) — under Apache 2.0, one of the most permissive open licenses. It's small enough to run on a single 24GB gaming GPU on your own machine. Meta also said it will "soon" release the weights of a bigger model, Muse Spark 1.2. Ethan Mollick called Spark the best non-Chinese open-weights model in a year; Simon Willison already had it running locally.
What exactly is an "open-weights" model?
When a company "opens the weights," it publishes the actual trained numbers (the billions of parameters) that make the model work — so anyone can download it, run it on their own hardware for free, inspect it, and fine-tune it. That's different from ChatGPT or Claude, which are "closed" — you can only rent access through their servers. Apache 2.0 is the key detail here: it lets you use the model commercially with almost no strings attached. For the past year the strongest open models mostly came out of China; Meta putting a capable agentic model back on the table shifts that balance.
What's the community saying?
- Local agents are the point — the excitement isn't raw benchmarks but that a genuinely agent-capable model now fits on consumer hardware, so you can build private, offline agents.
- Meta is "back" — many read this as Meta publicly recommitting to open weights after a quiet year, reframing the closed-vs-open debate.
- Still behind the frontier — Mollick and others temper it: Spark is good but trails both the closed frontier and the best Chinese open models, so "best non-Chinese" is doing real work in that sentence.