Benchmarks
How many bots a Paper 26.2 server can take, what they cost, and how the bots did in our fight tests.
We tested on a laptop with an Intel i7-14700HX, so a typical server host will handle fewer bots. The numbers still give you a good idea of how the cost grows. Short version: the server stayed at 20 TPS with around 300 bots fighting, and nothing broke or leaked in a 10 minute fight with bots dying and respawning the whole time.
Setup
| Date | 26 October 2026 |
|---|---|
| Server | Paper 26.2 build 134, Java 25.0.3, 4 GB of memory (6 GB for the runs over 250 bots) |
| PvpBot | The current release |
| World | Flat, no mobs, view and simulation distance 6, hard difficulty, no players online |
| Plugins | Only PvpBot |
| Machine | Intel Core i7-14700HX laptop, 32 GB RAM, Windows 11 |
Tick times come from Paper's /mspt, averaged over 20 seconds after the bots had been running for 30.
Memory is what was still in use after a full garbage collection. A tick has 50 ms to work with; go over that and TPS
drops below 20.
Load
"Idle" is bots just standing around with nobody to fight. "Brawl" is every bot fighting the ones near it, with Resistance V on everyone so the count doesn't drop. They still swing, hit and take knockback like normal.
| Bots | Idle: MSPT | Idle: worst tick | Idle: memory | Brawl: MSPT | Brawl: worst tick | Brawl: memory | TPS |
|---|---|---|---|---|---|---|---|
| 0 | 0.2 ms | 10.7 ms | 291 MB | 0.3 ms | 10.6 ms | 302 MB | 20 |
| 10 | 1.1 ms | 3.3 ms | 329 MB | 1.1 ms | 2.7 ms | 305 MB | 20 |
| 25 | 1.3 ms | 2.7 ms | 355 MB | 1.6 ms | 2.7 ms | 339 MB | 20 |
| 50 | 2.4 ms | 3.9 ms | 414 MB | 3.0 ms | 4.4 ms | 421 MB | 20 |
| 75 | 4.2 ms | 5.6 ms | 478 MB | 4.9 ms | 11.9 ms | 493 MB | 20 |
| 100 | 6.1 ms | 11.0 ms | 556 MB | 7.0 ms | 20.1 ms | 595 MB | 20 |
| 150 | 11.4 ms | 21.5 ms | 729 MB | 13.5 ms | 24.6 ms | 799 MB | 20 |
| 200 | 19.1 ms | 29.4 ms | 944 MB | 22.2 ms | 25.5 ms | 1045 MB | 20 |
| 250 | 31.8 ms | 46.3 ms | 1359 MB | 20 | |||
| 300 | 48.8 ms | 59.9 ms | 1679 MB | 20 | |||
| 350 | 68.8 ms | 83.2 ms | 1618 MB | 14.5 |
Each bot costs roughly 0.07 ms per tick and 3 MB of memory, and fighting is only 10 to 20% more expensive than standing still. That holds up to about 200 bots. After that every bot has more bots around it to keep track of, so the cost per bot goes up. On this machine 300 fighting bots was the limit before TPS started dropping. Memory was never the problem.
Long fight
30 bots fighting for 10 minutes with nothing keeping them alive. PvpBot removes a bot when it dies, so every 10 seconds we spawned more to get back to 30. That way bots were joining and dying the whole time, which is usually where memory leaks show up.
| Minute | MSPT (1 min avg) | Worst tick | Bots online | Memory | Deaths so far | Spawned so far | Errors |
|---|---|---|---|---|---|---|---|
| 1 | 3.1 ms | 156 ms | 22 | 371 MB | 29 | 50 | 0 |
| 2 | 2.2 ms | 138 ms | 29 | 375 MB | 59 | 87 | 0 |
| 3 | 2.3 ms | 146 ms | 28 | 374 MB | 97 | 121 | 0 |
| 4 | 2.1 ms | 133 ms | 26 | 370 MB | 126 | 150 | 0 |
| 5 | 3.6 ms | 416 ms | 27 | 413 MB | 159 | 185 | 0 |
| 6 | 2.1 ms | 118 ms | 24 | 382 MB | 189 | 212 | 0 |
| 7 | 2.1 ms | 145 ms | 26 | 384 MB | 218 | 242 | 0 |
| 8 | 2.2 ms | 152 ms | 28 | 392 MB | 252 | 277 | 0 |
| 9 | 2.0 ms | 131 ms | 28 | 391 MB | 276 | 302 | 0 |
| 10 | 2.1 ms | 146 ms | 25 | 384 MB | 308 | 332 | 0 |
No errors at all, and memory stayed between 370 and 410 MB from start to finish even with 332 bots coming and going. The high worst-tick numbers are from spawning: each top-up added up to 10 bots at once, and a batch like that takes a tick of 120 to 150 ms (one took 416). The average stayed around 2 ms, and without any spawning going on the worst tick at 25 bots is under 3 ms. If you're putting a lot of bots on a live server, spawn them a few at a time.
Fight tests
These are four fights between bots that we run before every combat change, with a minimum each one has to hit. Default config.
| Fight | Result | Needed | |
|---|---|---|---|
| Two spear and mace bots, 45 s | 3 smashes landed | 2 | Pass |
| Sword and shield vs sword and shield, 45 s | 29 sword hits, 21% of them blocked. The bots had their shields up 63% and 72% of the time when close | 10 hits, 20% blocked | Pass |
| Sword and axe vs someone holding a shield up the whole time, 45 s | 8 shield breaks, only 2 sword swings wasted on the shield, back to the sword half a tick after a break, about 7 hits in the 5 seconds after each break | 4 breaks, 4 wasted swings at most, back within 4 ticks, 4.5 hits | Pass |
| Sword and shield bot walking up to an archer 22 blocks away, 40 s | 22 of the 23 arrows that reached it blocked (96%), then 16 hits on the archer | Half blocked, 5 hits | Pass |
Every fight also checks that no bot froze up (close, ready to swing and not swinging for 3 seconds) and that PvpBot didn't throw any errors. None did.
How many bots should I run?
Up to 50 you won't notice them, that's about 3 ms a tick. 100 is around 7 ms and fine on most servers with players on. 200 is around 20 ms, which works if the rest of your server is light. Past 250 you're close to the limit even on a fast CPU, and most hosts are slower than this laptop. Plan for about 3 MB of extra memory per bot.
Running them yourself
The scripts are in PvpBot's source:
python pvpbot-paper/bench/bench.py # load: idle and brawl, 10 to 200 bots
python pvpbot-paper/bench/bench.py --only brawl --steps 250 300 350 400 --xmx 6G
python pvpbot-paper/bench/soak.py --bots 30 --minutes 10
python pvpbot-paper/fighttest/fighttest.py # fight tests
Each one sets up its own test server from an unpacked Paper install (PAPER_HOME), needs an
eula.txt you've accepted (EULA_FILE), and saves its results in .cache/.