Meta just released Glimmer, a new 30 billion parameter model built specifically for agentic workloads. According to Meta, it’s designed for function calling, coding, multi-step execution, and working inside existing agent frameworks, and it can run locally on a Mac or PC with a single consumer GPU. If you’ve got a MacBook Pro at home, you could be running an agentic model on device right now.
Meta deciding to put their chips behind a small, capable, locally runnable model is interesting on its own. It gets a lot more interesting once you look at Meta’s business model.
Meta isn’t selling you the model
In its most recent quarterly filing, Meta reported $56.3 billion in revenue. About $55 billion of that (roughly 98%) came from advertising. Meta’s core business is not selling access to AI models. It’s selling ads across its family of apps like Instagram, WhatsApp, and Facebook. That context makes Glimmer look a lot like a strategy a lot of us have seen before: Commoditize your Complements.
If you’re not familiar with this concept, the idea is simple. When your business depends on something you don’t directly monetize, you benefit when that thing gets cheaper and more widely available. In this situation, the “something” is agentic model capability: models that can call tools, execute multi-step tasks, and increasingly do useful work on their own. Meta is making that capability easier to run locally and deploy across infrastructure it doesn’t need to own.
To be clear, Meta hasn’t said it released Glimmer specifically to commoditize the model layer. That’s just my read. But the pattern holds regardless of intent: Meta makes nearly all of its money on advertising, but it keeps releasing models that push the cost and accessibility of AI capability down.
When the model gets cheap, the interesting question moves
As the trend continues, there’s a second order effect worth paying attention to across the agent ecosystem. When the model itself gets cheaper, simply having a model that can call a tool stops being a differentiator. That will be (if it is not already) table stakes.
The harder question becomes what happens when you actually let that model act.
Can it send an email reliably? Can it access Salesforce without messing up your system? Can it modify a production system without losing data? Whose identity is it using when it takes these actions? What permissions does it actually have? What policy governs the action, and what requires a human to sign off first?
None of those are model questions. They’re execution and governance questions, and they don’t get easier just because the model got smaller and cheaper to run. If anything, they get more urgent, because now more agents, on more infrastructure, are being tasked with taking action.
This is exactly the layer Arcade operates on
That’s why Glimmer’s release is such an interesting data point for us here at Arcade.dev. As the actions runtime for enterprise AI agents, we’re not in the business of making models smarter. We’re in the business of what happens the moment a model tries to act: enforcing whose identity an agent is using, executing that action reliably across the systems it touches, and governing what happened after the fact, in one place, with an audit trail.
Agentic intelligence is getting cheaper and easier to deploy. Glimmer is one more proof point in that direction, and it won’t be the last. As that trend continues, the infrastructure that matters most may increasingly have less to do with giving an agent intelligence, and a lot more to do with controlling what happens once that intelligence is allowed to act.
That’s the layer we work on every day at Arcade.dev. If you’re building agents, and Glimmer or any other model has made it easier to spin one up, you still need to govern what it’s actually allowed to touch. Get started at arcade.dev and connect it to thousands of tools with identity and policy already built in.
Go build something cool with Glimmer and let us know how it went!
