AI Pulse · Frontier Labs Watch

Meta just put a 30-billion-parameter model called Muse Glimmer into the wild under an Apache 2.0 license. It’s distilled from a larger, closed model they call Muse Spark, and it’s been quantized to run on under 20GB of RAM. That means you can run it locally, on hardware you already own, without needing a cloud bill or a data center.

The target use is coding agents. Think tools that write, debug, or refactor code on your machine, not in someone else’s cloud. Meta is betting that developers will build on Glimmer, get comfortable with the architecture, and eventually pull more of Meta’s stack into production. It’s a wedge strategy. Google and OpenAI keep their best models closed. Meta is seeding mindshare by giving away capable models, keeping the larger teacher model proprietary but letting the smaller one run free. They want you hooked on the ecosystem, not just the model.

For a business owner, this matters if you’re already running or considering local AI tools. A 30B model that fits in 20GB is suddenly viable for a lot of internal use cases, especially if you’re wary of sending proprietary code or data to a third party API. You get more control, lower recurring costs, and no rate limits. The tradeoff is you need someone who can deploy and tune it. This is exactly the kind of capability we build into an AI command centre, where the infrastructure and the workflows sit inside your environment, not rented by the token.

The broader pattern is clear. Open-weight releases are no longer just goodwill. They’re competitive positioning. If you’re planning your AI stack for the next 18 months, factor in that the line between “open” and “closed” is now a deliberate business decision by the labs, not a technical one.