Melee bots are getting really good

…and what we can learn from them

Melee (2001) runs at 60 frames a second and never waits. I put every kind of bot on one scale, plus LLMs, our own bot, a simulator and an AI commentary booth, in two days.

Mateo Amado · 2026-11-14 · melee.m19182.dev

A bench game redrawn from its replay: two Foxes on Final Destination

Ten years of bots, and why now

2001
Built-in CPUs, levels 1–9
2015
SmashBot: hand-written, frame-perfect
2017
Phillip: pure RL, beat pros at 2-frame reactions
2018–22
Slippi: replays, netplay, ranked. Data to learn from
2024
Phillip at human speed: lost to Zain, beat Moky
2025
HAL clones pro Fox for ~$5; MIMIC, per character
1, 17 Oct
Top players beat a Phillip crew on stream
Oct
Faynt: 240 of 244 vs Phillip. HAL scaled: top 1%
Oct
stomped.dev, a ranked ladder for bots. A ~100× simulator
11 Nov
Reyn: a bot you steer with English commands
Strong bots you can download, a language interface, a fast simulator, a bot ladder: all in the last six weeks.

How the best bots work

Game memory in, a full controller out, every frame. Copy humans, then self-play: AlphaGo’s recipe, minus the search.

Faynt · 10M

Transformer, 5 × 384, 256-frame cache. Cloned from 840k ranked games → “winner curriculum” → self-play RL in Fox mirrors. Zero delay.

Phillip · slippi-ai

Recurrent net; picks buttons (728 ways), then the stick (85). Cloned, then self-play. Sees the game 18–24 frames late, like a human.

HAL · 247M

Transformer, 256 frames of context, 1.3M games, 16 h on one B200. Predicts 4 frames of inputs, replans every 2. Pure cloning.

Reyn · Faynt + commands

Not a VLA: no LLM, no pixels. Trained on replays labelled with commands, then RL. A small router maps your words onto a fixed menu.
10M–247M
parameters: tiny next to an LLM
2–8 ms
per frame, on one home GPU
2×
HAL: a frame of reaction delay hurts twice as much as a frame of acting blind
Real-time board with 95% intervals

One scale for everyone

CPUs 1–9, three scripted bots, five learned bots and LLMs in five roles. Fox vs Fox on my home server, one Elo fit, CPU 9 = 1500. A scheduler picked the games: 1,573 overnight.

  • Learned bots are far ahead. Faynt never lost: 390–0. Then Phillip gm (2488) and HAL (2234)
  • CPU levels barely matter in a Fox mirror: level 1 (1250) beats level 5 (1194). The CPU Fox jumps off the stage by itself
  • LLMs span the whole board, depending on their job

Where the LLMs landed

Playing: speed beats smarts

1692 → 1114
GPT-6.1 Sol beats CPU 9 when the game waits for it. Live, 95% of its replies arrive too late, and every LLM sits below a level 1 CPU.

Coaching: two words win

“Don't die.”
Three of four LLMs coaching Reyn beat a random coach, by about 170 points. Holding “Don't die.” all game, with no coach, rates 2452: above every LLM coach.

Coding: a CPU 9 in 30 min

1663
Sonnet 5.5 wrote a 135-line bot just above the level 9 CPU. Opus's 303-line one did worse. One run each, so it ranks nothing yet.

A website to watch it all

The Bench page of the website
  • Bench: the board, a page per entrant with record and sample decisions
  • All matches: every game on the server, drawn live in 2D
  • Launch: any matchup, queued from the browser
  • Games: every replay, with the AI's decisions on a timeline
  • Live: the real game streamed, with what each AI is thinking
Public at melee.m19182.dev, dressed like Melee's own menus

An AI commentary booth

  • Two casters: Charlie on play-by-play (“loses his mind on big moments”) and Laura on colour (“dry wit, teases Charlie”)
  • A fast LLM watches the game's events (combos, stocks, self-destructs) and writes a line every few seconds
  • ElevenLabs voices them, only while someone is listening
Charlie: HAL MASTER's Fox is in the air and Faynt's Marth is waiting for him!
Laura: Waiting is the right word. Faynt doesn't chase, Charlie, he collects.

Our own bot: m19182-yoshi

Training page: loss and validation accuracy of the 20M model
23,000
ranked online Yoshi games to learn from
  • A small transformer that copies human play, the way HAL did: 4 s of game in, the next controller input out
  • 6M: 0 wins in 12 against CPU 9. 20M: 1 win, twice the stocks taken
  • Next: self-play, where the published bots got strong

A simulator, ~100× faster

287 / 300
replays reproduced frame for frame (the other 13 were unfinished files)
~100×
faster than the emulator
  • The community decompiled Melee back into C; melee-sim-light runs it with no emulator
  • I hooked every player into it, learned bots included
  • Save any moment, try thousands of inputs, keep the best: tool-assisted superplay, and training on far more games than humans play
  • Catch: no built-in CPUs

Play the bots yourself

  • On my laptop with a controller, after this: Phillip at any tier, Faynt, Reyn, or our m19182-yoshi
  • From home, over Slippi direct connect
  • The board has a slot waiting for humans
Write-up, boards, replays and live games
melee.m19182.dev/writeup
0:00