
Model collection · Showcase · Commands · Inference quickstart
Reyn is a language-steerable model for Super Smash Bros. Melee, supporting all 26 characters. Use natural-language instructions to steer play style, movement and combat. Pass in the model inputs and an instruction to get the next predicted action. The same loaded weights handle every character.
Online play: Please do not use these models in online matches without explicit permission from the other participants and the relevant platform or event organizers. Follow the platform's rules.
81.8% of test cases met their command-following target
1,368 / 1,672 cases across the original-command and main special-move suites.
Language steering in action
Watch the orange Fox use Fire Fox, reach the top platform, grab and fire lasers from a distance. Each panel starts with two seconds of unsteered play. The bar turns green when the language instruction activates, with Reyn and the instruction labelled above Fox.
Each panel is a continuous clip from a separate game, shown at normal speed.
At a glance
| Field | Details |
|---|---|
| Characters | All 26 Melee characters |
| Stages | Six stages, with frozen Pokémon Stadium |
| Commands | Play style, movement, platforms, attacks, defense and special moves |
| Input → output | Structured tensors → predicted action, labels and logits |
| Runtime | PyTorch · weights stay fixed during inference |
Use the model
Install the inference package, then download this complete repository:
python -m pip install huggingface_hub
hf download frisson-labs/Reyn-26C \
--local-dir model-files
from reyn import Reyn
model = Reyn.load(
"model-files", character="peach"
)
inputs = model.example_input()
prediction = model.predict(
inputs, phrase="pull a turnip"
)
print(prediction.action)
The example uses synthetic tensors. Replace the values with your model inputs, keeping the field names, shapes and dtypes. The download includes four matching model files and the required checksums.sha256. Keep them together; the loader verifies them before use.
An accepted command stays active for later predictions. model.say(phrase) returns the router's reply; an "abstain" reply leaves the previous command active. model.clear() cancels the command directly. model.reset() also clears prediction history. model.set_context("fox") switches character and clears the command and history without reloading weights.
Commands
Use language steering to shape play style, movement and combat. Ask Reyn to use platforms, attack, defend or look for a combo.
| Category | Example instructions |
|---|---|
| Play style | Be more aggressive. · Play defensive. · Play safe. |
| Movement | Jump from the ground. · Dash dance. |
| Position | Go to the left side. · Keep to the center of the stage. |
| Platforms | Go to the top platform. · Stay on the main floor. |
| Attacks and grabs | Use forward smash. · Attempt a grab. |
| Defense | Put your shield up. · Choose a spot dodge. |
| Combos | Do a three-hit combo. · Finish your combo with an up-air. |
| Ledges and projectiles | Take the ledge. · Keep your distance and shoot lasers. |
| Character specials | pull a turnip · fire fox · mario fireball |
The full command reference covers every supported command meaning, plus all 104 character-specific special moves and how to change or clear a command. Choose moves and destinations that fit the character and stage.
The special-move CSV gives two example instructions per move and marks required character cues. The neural router reads the phrase as written; character and stage inputs do not rewrite it.
Results
Combined result: 81.8% (1,368/1,672 cases). This combines the 386/424 original-suite cases and 982/1,248 main special-move cases reported below.
On the general-command benchmark, 386/424 cases met at least the looser target (91.0%): 267 met the stricter target, 119 met only the looser target and 38 met neither. The looser target allows a weaker threshold or a missed extra check from the stricter target. This test covers Fox, Falco, Marth and Mario across six stages. The same instruction can appear in more than one condition, so these are test cases, not distinct commands or match wins.
The special-move tests cover all 26 characters against level-9 Fox on six stages. Useful execution reached 974/1,248 (78.0%) for short phrases and 982/1,248 (78.7%) for longer phrases. Each suite repeats two phrasings of 104 special-move intents on six stages. A useful result requires the requested move and its applicable effect. Unknown and censored opportunities stay in the totals.
The short test uses 208 generated intent/phrase rows with 186 distinct strings. These results measure command following and special moves; they do not establish all-character general-command coverage, speech recognition, response latency or fighting strength.
Stages are Final Destination, Battlefield, Dream Land, Yoshi's Story, Fountain of Dreams and frozen Pokémon Stadium. 2v2 and play with a human partner use separate releases. Wombo Combo is a separate experiment made for fun.
License
Frisson Labs' inference code and model weights are released under the MIT License. See the third-party notices for upstream attributions.
The online-play warning is guidance; the MIT license terms are unchanged.
- Downloads last month
- 59