Llama-4 fine-tune shipping to 800M MAU: which license clause must the team re-read?

Short answer·Medium·4.0 · 0·~3 min·Relevant atMeta Mistral AI

Attempt it

A product team is about to ship a Llama-4 fine-tune as part of a consumer feature inside an app with 800 million monthly active users. They've assumed the Llama community license is "basically open" and have not requested anything from Meta. Identify the specific clause they need to re-read, what threshold it sets, why it applies to a fine-tuned derivative as much as to the base model, and what they need to do BEFORE shipping.

Free · 2 AI evals / day

**The Llama Community License** is not a permissive open-source licence like Apache 2 or MIT. It allows free use under most conditions, but it carries one specific commercial threshold the team has missed. The clause is the monthly active user threshold. If the products or services you make available exceed 700 million monthly active users in the preceding calendar month, the community licence stops granting you rights. You must request a separate commercial licence from Meta before you can use Llama or any derivative of it. Three facts the team has to internalise. First, fine-tuning does not reset the licence. A LoRA adapter or a merged fine-tune is a derivative of the base weights, and the licence travels with the derivative. The fine-tuned model is still bound by the same threshold. Second, the threshold is measured against the product surface, not against the AI feature. An 800 million MAU consumer app blows past 700 million the moment the model ships inside it, even if only 5 percent of users ever touch the AI feature. Third, the required action is operational. Contact Meta, request a commercial licence, and do not ship until you have one. If Meta declines or does not reply in time, the alternatives are to swap to a base model under a permissive licence like Mistral or Qwen, or to restrict the deployment surface to one under the threshold.

**The clause is the monthly active user threshold in the Llama Community License.** It sets a numeric boundary: if the products or services made available by the licensee, its affiliates, or its corporate group exceed 700 million monthly active users in the preceding calendar month at the time of release, the standard community licence does not grant rights to use the Llama model or its derivatives. A separate commercial licence from Meta is required, and Meta has sole discretion over whether to grant it. Three implications the team needs to act on. First, a fine-tune is a derivative work. LoRA adapters, fully merged fine-tunes, distilled checkpoints, all inherit the upstream licence terms. Outputs generated from a Llama-based system inherit them too. Training your own adapter does not produce a fresh licence-free model. Second, the threshold is measured against the deployed surface, not against the AI feature usage. An 800 million MAU consumer app crosses 700 million by definition, even if only a small minority of users interact with the AI capability. The licence wording covers products or services made available, which captures the whole app. Third, the operational sequence is: contact Meta, request the commercial licence, wait for it to be granted, then ship. If Meta declines or does not respond on the timeline you need, the workable alternatives are to swap to an Apache 2 or MIT licensed base like Mistral or Qwen variants, or to restrict the Llama-based deployment to a sub-surface that is genuinely below the threshold. Doing neither and shipping anyway is a contractual breach with material exposure.

Quick reference

TL;DR

The Llama Community License triggers a separate Meta commercial-license requirement above 700 million MAU on the deployed product. Fine-tuning does not reset the clock, the derivative inherits the same threshold.

Memory aid

Easy to grasp

Imagine borrowing a recipe from a chef who lets anyone use it, except for restaurants serving more than seven hundred million meals a month. Those huge restaurants have to call the chef first and arrange a separate deal before they can serve any dish that uses the recipe. Tweaking the recipe with extra spices does not change anything, because the new dish still uses the original. The size of the restaurant you serve in is what triggers the call, not whether the recipe section of the menu is busy. So if your restaurant chain has over seven hundred million monthly customers, you make the call before you open the kitchen.

Key concepts

Model licensing Llama models Commercial license Compliance

Detailed answer & concept explanation

~8 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Quick scan · 2–4 min

Everything important, quickly.

5 min: identify the 700M MAU clause + why the deployment surface counts + why fine-tuning does not reset the licence + the corporate-group language + the operational sequence to request the commercial licence + the swap to permissive fallback + why the licence is not equivalent to open source.

Real-world examples

Real products, models, and research that use this idea.

Meta's Llama 4 Maverick released under the Llama Community License in 2026, with the same MAU threshold structure as Llama 2 and Llama 3.
Mistral's Apache 2 licensed open-weight models like Mistral Small 3 are commonly chosen by large-MAU products specifically to avoid this constraint.
Qwen 3.5 under Apache 2 from Alibaba is another common Llama swap target for products that cross the threshold.
Several large consumer apps in 2026 have publicly switched away from Llama bases or negotiated separate commercial deals with Meta when their MAU crossed into the gated zone.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you advise a company whose MAU is currently 500 million but projected to cross 700 million within the next year?

QDoes the licence treat synthetic data generated by Llama and used to fine-tune a different model as a derivative?

QWhat is the difference between the Llama Community License and a genuinely permissive licence like Apache 2?

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Believing that fine-tuning produces a fresh license-free model. The derivative inherits the upstream terms, including the MAU threshold, the LoRA training step changes nothing about that.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

References

Primary sources. Skim if you want the original framing.