On Tuesday, Meta released Muse Spark 1.3 — its fourth model in five months. The public gets “xhigh,” which scores 61 on the Artificial Analysis Intelligence Index, tied with GPT-5.6 Sol and Grok 4.6. Partners get “max,” which scores 62, behind only Anthropic’s two flagships. Open weights, Mark Zuckerberg promised, are “coming soon.”

The conventional read is that Meta is finally getting serious about AI — closing the gap, spending the money, competing with Anthropic and OpenAI. The open-source purists are upset, but that’s just sentiment. The real story is the tiering.

The Frontier Is Now a Waiting Room

Meta has quietly redefined “frontier” to mean “what you can’t have yet.” The max variant — the one that actually competes with Claude Fable 5.1 — is in “limited partner preview.” The public gets xhigh, which is four points behind. The open weights are “coming soon,” which is a promise, not a delivery.

The pivot started in June 2025, when Meta reorganized its AI labs after critics alleged that Llama 4’s training data had been contaminated with benchmark answers. The $14.3 billion Scale AI stake and Alexandr Wang’s arrival as chief AI officer followed. The open-weights era ended quietly, without a press release.

This is a retention tool. The open-weights community that built on Llama is being strung along with a promise that, when it arrives, will be obsolete on arrival. Four releases in five months means the model you can actually use is always the one that’s already been superseded.

One developer who built a production system on Llama 3 put it in a Slack DM: “I’m not mad they went proprietary. I’m mad they’re pretending they didn’t.”

The Cadence Is the Product

Four releases in five months. That’s not a product strategy; that’s a treadmill. The scores jumped 8 points from July to August, then 4 more. The “frontier” is now a moving target that Meta controls the pace of. The public model is always one step behind the partner model. The open weights are always “coming soon.”

This is the same playbook as everyone else — release a “frontier” model to partners, release a slightly worse one to the public, promise open weights later. The difference is that Meta used to be the open-weights company. Now it’s just another tiered-access company.

The “almost too cheap to meter” line from Wang is telling. Cheap to meter for whom? The partners who get max? The public who gets xhigh? The developers waiting for open weights? The answer is: cheap for Meta, because the tiering means the frontier is always just out of reach.

And the next model is already teed up — Zuckerberg’s tweet ended with “Next up 🍉.” The watermelon is coming. The treadmill doesn’t stop.

What the Open-Weights Promise Is Actually For

The open-weights promise is doing a lot of work. It keeps the open-source community’s goodwill without actually ceding the frontier. It gives developers a reason to stay on Meta’s platform while the proprietary model gets better every month. It’s a retention tool, not a strategy.

The Hacker News thread on Tuesday’s release — 653 points, 427 comments — is a decent proxy for the open-weights community’s mood. They’re not angry. They’re waiting. And waiting is exactly what Meta wants.

The developers who built on Llama are now being told to wait. The “frontier” is defined by what partners get, not what the public gets. And the cadence means the open-weights release, when it comes, will be a snapshot of a model that’s already been superseded.

That’s the point. The promise isn’t a commitment; it’s a leash. And the developers who built on Llama are wearing it.

Sources