| # Model Card | |
| - Source: [https://arxiv.org/abs/2509.02046](https://arxiv.org/abs/2509.02046) | |
| - Optimizer: `cautious` | |
| - Model size: `520m` | |
| - Data size: `21B` | |
| ## Best configuration | |
| | Hyperparameter | Value | | |
| |---|---| | |
| | beta1 | `0.98` | | |
| | beta2 | `0.98` | | |
| | epsilon | `9.999999999999999e-26` | | |
| | learning_rate | `0.008` | | |
| | max_grad_norm | `1` | | |
| | min_lr_ratio | `0` | | |
| | train_batch_size | `256` | | |
| | warmup | `2000` | | |
| | weight_decay | `0.1` | | |