Meta Drops the User Cap: Llama 4 Now Fully Commercial

Meta has updated the Llama 4 license to remove the clause that previously barred companies with more than 700 million monthly active users from deploying the models commercially. The change applies to both Scout, a 17-billion-parameter dense model, and Maverick, a 400-billion-parameter mixture-of-experts architecture. Both sets of weights are downloadable directly from Hugging Face under the new terms.

The 700-million-user cap had been a notable friction point for large-scale commercial deployment, effectively excluding the biggest potential adopters — cloud providers, large enterprises, and consumer platforms — from building on Llama 4 without separate negotiation with Meta. Removing it aligns Llama 4 more closely with permissive open-weight licenses, though it still isn't OSI-approved open source; the Llama license retains restrictions on using the models to train competing foundation models.

Scout at 17B is positioned as an efficient, deployable model suitable for on-device or cost-constrained inference, while Maverick's MoE architecture provides high capability at reduced active-parameter cost per token. The practical implication is that any company can now integrate either model into production products, fine-tune on proprietary data, and serve users at scale without hitting a licensing ceiling or negotiating enterprise terms with Meta.

This move continues Meta's pattern of using open-weight model releases as a strategic distribution play — lowering the barrier to Llama adoption drives ecosystem lock-in through tooling, fine-tunes, and developer familiarity, even if Meta itself doesn't directly monetize the weights. The question now is whether the remaining license restrictions, particularly the prohibition on training competing models, will limit adoption among the labs and cloud providers most capable of deploying at this scale.

Panel Takes

The Builder

Developer Perspective

“The primitive here is simple: pull weights from Hugging Face, run inference, ship product, no license audit required at any scale. That's a real unblocking for teams who were previously doing legal gymnastics around the 700M-user clause before they even had 700 users. The remaining restriction — no training competing foundation models — is targeted enough that it won't affect 99% of production use cases, and the MoE architecture on Maverick means you're paying active-parameter costs, not full 400B costs per token, which changes the infrastructure math considerably.”

The Skeptic

Reality Check

“The 700M-user cap removal sounds dramatic but in practice the companies it blocked were the ones most capable of negotiating custom terms anyway — this is a cleanup, not a revolution. The more interesting restriction that stays is the prohibition on training competing foundation models, which means the hyperscalers building their own model stacks still can't use Llama 4 as a training signal, and that's the population Meta actually cares about closing off. What kills this positioning in 18 months isn't a competitor — it's Meta itself shipping Llama 5 with better benchmarks and making Llama 4 irrelevant before anyone has built serious production infrastructure on it.”

The Futurist

Big Picture

“The thesis Meta is betting on: open-weight models become the default infrastructure layer for enterprise AI, and whoever's weights are most widely embedded when standardization happens wins the ecosystem the way Linux won servers. Removing the user cap is less about revenue and more about ensuring Llama is the model in the fine-tune, the edge deployment, the internal tool — so that switching costs accumulate in Meta's favor without Meta charging for them. The dependency to watch is whether the 'no competing model training' clause holds up legally and competitively as synthetic data pipelines make model-on-model training increasingly hard to audit.”

The Founder

Business & Market

“Meta isn't monetizing these weights directly, so the right question is what they're buying with the giveaway — and the answer is ecosystem gravity: every startup that builds on Llama is a distribution node, every fine-tune is a switching cost, every developer who learns Llama tooling is less likely to retrain on a competitor's base. The business risk for startups building Llama 4 wrappers is unchanged: if Meta decides to productize anything you're building on top of Scout or Maverick, your moat is whatever data and workflow integration you've accumulated, not the model itself. The license change makes Llama 4 a better foundation to build on, but it doesn't change the fundamental defensibility question for anyone downstream.”

Panel Takes

Bookmarks