AFM-4.5B
Arcee AI · 2025-07-29 · 4.6B parameters
AFM-4.5B is Arcee AI's open-weight instruction model with 4,619,189,760 checkpoint parameters. Arcee reports base-model pretraining and mid-training exposure only as a rounded 8T-token figure, followed by SFT and reinforcement-learning post-training with undisclosed token counts; its context was extended from 4,096 to 65,536 tokens using YaRN-scaled RoPE, and the current Hugging Face release is Apache-2.0 licensed.
Benchmark scores
| Benchmark | Score |
|---|---|
| SnakeBench Average Apples | 0.2 |
| SnakeBench Best Apples | 2.0 |
| SnakeBench Rating | 13.3 |
| SnakeBench Total Apples | 9.0 |
| SnakeBench Win Rate | 39.5 |