Most people pick their server software by asking which is fastest. That is the wrong first question. The right one is what you want to be able to install, because that decision is close to irreversible and the performance differences are smaller than the compatibility ones.
Four realistic options, and what each one actually costs you.
Vanilla
The server jar Mojang publishes. Runs the game exactly as designed, accepts no plugins and no mods, and is the only option guaranteed to work on day one of every update.
Costs: no permissions system, no land claims, no anti-grief, no economy, no admin tooling beyond operator commands. Performance is the worst of the four under load, because none of the optimisation work of the last decade is in it.
Choose it when: five friends, a whitelist, and nobody wants to administer anything. That is a real and common case, and vanilla serves it better than the alternatives because there is nothing to maintain.
Paper
The mainstream plugin server. Runs Bukkit and Spigot plugins, which is the largest ecosystem of server-side functionality that exists – permissions, claims, chat management, minigames, economies, moderation tooling.
Costs: no client mods. Paper cannot run Fabric or NeoForge mods, and anything that adds blocks, items or mobs is out. It also changes some vanilla behaviour in the name of performance, which matters if your server runs technical farms – some redstone and mob mechanics differ subtly from vanilla, and the differences are configurable but you have to know to look.
Choose it when: the server is public or semi-public, you need to manage people rather than content, and the vanilla feature set is what you want to play.
Fabric
The lightweight mod loader, server side. Runs Fabric mods, which is where the performance stack lives – Lithium, C2ME, ScalableLux – and where most client-visible content mods have gone.
Costs: no Bukkit plugins, so administration is handled by mods that are generally less mature than the plugin equivalents. Every content mod has to be installed by every player, which is a real barrier for a public server and a non-issue for a private one.
Choose it when: you want the performance mods, you want content mods, or both, and your players will install a mod folder. The performance stack alone is a legitimate reason to pick Fabric for an otherwise vanilla server, since those mods are server-side and need nothing from players.
NeoForge
The heavier mod loader, and where the large tech and magic mods live. If your server is built around a modpack, the pack chose this for you and there is no decision to make.
Costs: slower to update after a drop than Fabric, heavier at runtime, and a smaller server-side performance mod selection. Version-locked to whatever your pack targets.
Choose it when: you are running a modpack, or a specific large mod only exists here. The loader comparison covers the client-side half of this.
The hybrid question
There are projects that run plugins and mods together. They work, and they are the least stable category of server software because they are reconciling two things that were not designed to coexist.
The honest advice is to avoid them unless you have a specific requirement that nothing else meets, and to expect update delays longer than either parent project. If the requirement is really “plugins for admin, mods for content”, check whether a Fabric equivalent of the admin plugin exists first – increasingly one does.
What actually changes performance
Switching from vanilla to Paper or to Fabric plus the performance stack is a real improvement, and it is also not the largest lever you have.
The larger ones, roughly in order: view distance and simulation distance, whether the world has been pre-generated, how many chunks are being force-loaded, and how much RAM the JVM is actually allowed to use. A badly configured Paper server loses to a well-configured vanilla one.
Hardware sits underneath all of it, and single-thread performance matters more than core count for everything except chunk generation – the reasoning is in the server specs breakdown.
Switching later
Vanilla to Paper or to a loader is generally safe – the world format is the same and both read it. Going back from Paper to vanilla is also fine.
What does not travel is anything the software added. Plugin data, claims, economies and mod-added blocks do not survive a switch away from the thing that created them. A world with modded blocks in it will load on vanilla with those blocks removed, permanently.
So the reversible decision is the loader, and the irreversible one is the content you let into the world. Decide the second one deliberately, take a backup before any switch, and treat the migration as its own job – the checklist is the same as a version migration.
Short version
- Small private group, no admin appetite: vanilla.
- Public server, people to manage: Paper.
- Performance mods, content mods, or a small SMP willing to install things: Fabric.
- Running a modpack: NeoForge, decided for you.
If you have not picked a host yet, that decision interacts with this one – some hosts make loader installs trivial and others do not. Choosing a host covers what to check.
Images: Dungeons and Taverns Pillager Outpost Overhaul project gallery.