Llama 4: Meta Finally Figured Out How to Make Money From Open-Source AI

Meta dropped Llama 4 this week, and the story isn't just about benchmark numbers. It's about two things that actually matter: how the model works under the hood, and how Meta has finally stopped pretending open-source AI doesn't need a business model.

The Sparsity Play

Llama 4's headline technical feature is sparsity-driven inference, which sounds like marketing speak until you look at what it actually does. The model is built on a Mixture-of-Experts, or MoE, architecture. Instead of activating all parameters for every token, the model routes each input to a small subset of expert sub-networks, typically two to four out of roughly 128 total.

This isn't a new idea, but Meta's implementation is notable for the scale at which it works. The result is a 40% reduction in floating-point operations per token, which translates directly to lower latency, lower energy use, and the ability to run a model this size on hardware that previously couldn't handle it. Only the activated experts need to sit in memory during inference, so VRAM requirements drop accordingly.

For developers, this is genuinely useful. A model that can match or beat GPT-4o on multimodal tasks while using significantly less compute per token changes the deployment math for a lot of teams. The Maverick variant, with 17 billion active parameters, scored 73.4% on the MMMU multimodal benchmark, which puts it ahead of GPT-4o at 69.1% and Gemini 2.5 Pro at 71.7%.

The caveat, as always with MoE models, is that the total parameter count is much larger than the active count. The full model still needs to be stored somewhere, and routing overhead can introduce complexity that doesn't exist in dense architectures. But for inference-heavy workloads, the tradeoff clearly makes sense.

The Licensing Pivot

The bigger shift, and the one that will shape the open-source AI landscape more than any architecture change, is Llama 4's new tiered license structure.

Llama 3 operated under a single Community License that allowed commercial use up to 700 million users, with an Acceptable Use Policy that prohibited a predictable list of harmful applications. It was functional, but it left a lot of ambiguity for companies operating at scale.

Llama 4 replaces that with two distinct tiers. The Community License remains free and covers research, academics, and limited commercial experiments under the same 700 million user threshold, with attribution and AUP requirements. The Enterprise License is new. It is paid, with fees tied to user base and revenue tier, and it adds explicit compliance obligations, custom usage clauses, and the option for dedicated support and SLAs.

This is Meta drawing a hard line between "open source for hobbyists and researchers" and "open source for businesses making real money." It is not an unreasonable line. Running a model at enterprise scale creates costs for Meta, both in terms of support burden and in the risk that misuse by a large platform could blow back on them politically.

But it does change the character of what Llama is. For the past two years, Llama has been the default choice for startups and developers who wanted a capable model without negotiating with a cloud API provider or paying per-token fees. Llama 4 is still that, up to a point. Past that point, it becomes a negotiated commercial relationship, much like what you would have with OpenAI or Anthropic.

The open-source purists will call this a retreat. It isn't. It is a clarification. Meta was never going to subsidize every company's AI stack indefinitely. The question was always when the free lunch would end, not if.

What This Means for the Ecosystem

The licensing change has immediate implications for the competitive landscape. Companies that have built products on top of Llama 3 and are now scaling past 700 million users face a decision: pay Meta for an enterprise license, migrate to a truly open-weight model like Mistral or Qwen, or switch to a closed API entirely.

That 700 million threshold sounds enormous, but it is not as high as it used to be. A consumer app with moderate global adoption can hit it quickly. Meta is effectively saying that if you are big enough to matter, they want a cut.

For smaller developers and researchers, nothing changes. Llama 4 is still the most capable freely available model for most tasks, and the MoE efficiency improvements make it more practical to self-host than ever. If you are running a side project or a small SaaS, you are probably not thinking about enterprise licensing yet.

For the rest of the industry, Meta's move validates what OpenAI and Anthropic have been saying all along: there is no sustainable future in giving away frontier models for free. The open-source community will keep building, but the economics of training models at this scale are forcing everyone, including Meta, to build revenue models around them.

Bottom Line

Llama 4 is a technically impressive release. The MoE architecture delivers real efficiency gains, and the benchmark numbers hold up against the closed competition. The licensing pivot is the more consequential story, though. It signals that the era of unlimited free Llama is over, and that Meta intends to capture value from the ecosystem it built.

That is not a criticism. It is just what happens when something becomes important enough to matter.