Meta open-sources a 30B model and teases a stronger one

Meta released Muse Glimmer on Monday, a 30-billion-parameter open-weight model built to run AI agents locally on consumer hardware. It ships under Apache 2.0, which means you can download it, modify it and use it commercially without asking anyone.

Glimmer is essentially an open version of Muse Spark, the closed model Meta launched in April. And the company says it plans to open the weights for Muse Spark 1.2 as well, its most advanced model, and that is the part that actually matters here.

That’s a different posture from everyone else at the frontier. OpenAI and Anthropic keep their strongest models closed, so a lab open-sourcing its best one is either a strategic bet or a concession, depending on who you ask.

Mark Zuckerberg made the case himself in a 6,500-word essay published alongside the release, arguing that powerful AI shouldn’t end up controlled by a handful of companies. ABC News read it as a manifesto, and it lands as a fairly direct swipe at his closed-source competitors.

A 30B model under Apache 2.0 is not charity. It is a bid to make Meta’s architecture the default thing developers build on.

The strategy behind a Meta open source model

Because that’s the strategic logic, and it’s worth being clear-eyed about it. Meta doesn’t sell model access, so it loses nothing by giving weights away. What it gains is an ecosystem: tooling, fine-tunes and developer habits that form around whatever runs well locally. The same logic shows up in how frontier pricing is moving, where the cheaper option keeps winning on volume.

The 30B size is the tell. That’s small enough to run on a good consumer GPU and large enough to be genuinely useful, which puts it exactly where hobbyists and startups build. CNBC framed the release as a swipe at OpenAI and Anthropic, and commercially that reading holds up.

It also fits a personal-computing pitch rather than a data-centre one. TechCrunch read Glimmer as a hint at Zuckerberg’s “personal intelligence” idea. Models that live on your device instead of behind someone’s API.

Zuckerberg also used the moment to criticise US restrictions that, in his telling, benefit foreign labs, per Yahoo Finance. That argument is doing double duty. It’s a policy position and a defence of shipping open weights while regulators are still deciding how they feel about that.

What a Meta open source model cannot undo

Which brings up the awkward part. Open weights can’t be recalled. Once they’re downloaded, safety mitigations can be fine-tuned away by anyone with a modest budget, and no amount of licence language changes that. We’ve covered where export controls sit on this, and the honest answer is that the policy hasn’t caught up.

Worth watching whether Spark 1.2 actually ships open. Announcing intent is free; releasing frontier weights is the commitment, and plans have quietly changed before.

The strategy behind releases like this is worth understanding, and our piece on why labs publish weights covers what open weights do and do not give you.

Converting that model work into revenue is the harder half, which our piece on Meta and enterprise sales takes up.

If it does land, the interesting question isn’t whether open models catch the frontier. It’s whether “good enough, free, and running on your own machine” turns out to matter more than the last few points on a benchmark. That is the same bet the Grok pricing story is making from a different direction.