Explore
our models
Compare the two model sizes and follow each from pretraining through supervised post-training to reinforcement learning.
Base: pretrained. Expert: supervised post-training. Arena: reinforcement-learning checkpoint. RL uses Fox mirrors; the same weights continue to support all 26 characters. Consult each model card for usage and licensing.
Two sizes.
One interface.
Both models consume structured game state and predict controller inputs. Training uses 256-frame windows.
Read the full training recipe| Configuration | 10M | 75M |
|---|---|---|
| Parameters | 10,163,629 | 75,305,709 |
| Transformer blocks | 5 | 11 |
| Hidden width | 384 | 768 |
| Query / KV heads | 6 / 2 | 12 / 3 |
| Context | 256 frames | 256 frames |