Frisson Labs releases language-controllable Melee model Reyn 26C
Understand this story
Frisson Labs releases Reyn 26C, a language-controllable Super Smash Bros. Melee agent built on its Faynt, covering all 26 characters. According to its published data, 81.8% of test cases met the instruction-following goal. Reyn 26C was trained on about 840,000 human replays, with imitation learning and RL fine-tuning, and about 2.2 million parameters participate in instruction-related training. Frisson Labs says character support remains uneven, and special-move commands for characters such as Mewtwo and Zelda are still unreliable. 2v2 team collaboration is currently a separate experimental branch, using only Fox, and is separate from the current Reyn 26C version. That branch has an 84.4% win rate against two Faynt setups, and teammates take 71%–74% less damage.
AI overview based on source reports · Updated 36 min ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIWe got AI to Wombo Combo – collaborative agents in Melee
The team behind Faynt, an open-weight Super Smash Bros. Melee model trained on roughly 840,000 human replays, introduces Reyn 26C, a language-steerable version that handles special-move requests across all 26 characters and met at least the looser command-following target in 386 of 424 test cases. A separate experimental Fox 2v2 branch, trained with full policy parameters rather than adapters, won 84.4% of 160 real games against two Faynts, and a play-time action mask using Fox's hitbox data cut teammate damage by 71–74%. A bonus Wombo Combo 2v1 experiment produced repeatable routines against an idle Bowser, but performance dropped sharply against opponents that could act.
Activity for this story
Not enough continuous observations to show a trend yet.