Llama 4 Maverick

Maverick is the mid-tier of the Llama 4 family — the size most self-hosting teams actually deploy, because it fits on hardware they can justify while remaining capable enough for production work.

If Behemoth is the statement of what open weights can do, Maverick is the one people run.

The practical sweet spot. It serves on a realistic GPU budget, handles the ordinary work of a production system — drafting, summarising, extraction, structured output — and does so without per-token cost.

Fine-tuning at this size is achievable rather than theoretical, which is where open weights pay off most: a mid-sized model tuned on your domain frequently beats a larger general model on your specific task, and costs less to run.

It needs more prompt engineering than a hosted flagship. Behaviour that a closed frontier model gets right by default has to be specified, and output validation matters more.

You are also responsible for everything: serving reliability, scaling, security patching, evaluation when you change anything. That work is invisible until it is urgent.

Maverick is the default choice for teams self-hosting seriously. Start here rather than at Behemoth — prove the workload works on hardware you can afford before scaling up, and consider fine-tuning before assuming you need a bigger model.