Remember when Meta said they were all-in on open source AI? Yeah, about that. The company is now preparing to release new models developed under Alexandr Wang’s leadership, and surprise—not all of them will be open source. Shocking, I know.
Let me be clear: I’m not against selective open sourcing. What bugs me is the messaging. Meta has spent months positioning itself as the champion of open AI development, the scrappy underdog fighting against OpenAI’s closed gardens and Anthropic’s safety theater. Now we’re learning that “open source” actually means “open source when we feel like it.”
What We Actually Know
The facts are thin, which tells you something right there. Meta plans to release open source versions of its next AI models—note the plural “versions,” not “models.” Some models will get the open treatment. Others won’t. The company frames this as a strategy to enhance collaboration and accelerate innovation, which is corporate speak for “we’ll open source what doesn’t hurt our competitive position.”
This is the first batch of models developed under Wang’s watch, so there’s genuine curiosity about what direction he’s taking things. But the selective approach undermines the whole narrative Meta has been pushing.
The Open Source Bait and Switch
Here’s what frustrates me as someone who actually tests these tools: Meta’s previous open releases like Llama have been genuinely useful for the developer community. They’ve enabled smaller companies and independent researchers to build things they couldn’t afford to create from scratch. That’s real value.
But now Meta wants credit for being “open” even as they cherry-pick which models to release. You can’t have it both ways. Either you’re committed to open development or you’re running a traditional tech company that occasionally throws the community some bones. Both are valid business strategies, but pick one and own it.
What This Means for Developers
If you’re building on Meta’s AI infrastructure, this should make you nervous. The company is clearly keeping its best stuff proprietary. The models they do open source will likely be the ones where they’ve already extracted maximum competitive advantage, or where open sourcing serves their platform strategy.
That’s not necessarily bad—you can still build useful applications on top of these releases. But go in with your eyes open. You’re not getting access to Meta’s crown jewels. You’re getting access to what Meta has decided won’t hurt them to share.
The Bigger Picture
This move reflects a broader trend in AI development. The early days of “let’s all share everything” are over. Companies have figured out that AI models are valuable intellectual property, and they’re getting more strategic about what they release.
Meta’s approach is actually more honest than some competitors who’ve quietly backed away from open commitments without saying anything. At least Meta is being upfront that not everything will be open. But that doesn’t mean we should applaud them for doing the bare minimum.
The real test will be what they actually release. If the open source versions are genuinely useful and not just marketing props, then fine. If they’re neutered versions that exist mainly so Meta can claim the “open source” label, that’s a different story.
My Take
I’ll reserve final judgment until we see what Meta actually ships. But based on the limited information available, this looks like a company trying to have its cake and eat it too. They want the PR benefits of being “open” without the competitive risks of actually opening everything up.
For developers and companies building on Meta’s AI: plan accordingly. Don’t assume you’ll get access to their best models. Build your strategy around the assumption that you’re working with their B-team releases, and be pleasantly surprised if you get more.
And Meta, if you’re listening: just be honest about what you’re doing. There’s no shame in keeping some models proprietary. But stop pretending you’re the open source heroes of AI when you’re clearly not.
🕒 Published: