AFM-4.5B

AFM-4.5B is Arcee's 4.5B instruction-tuned foundation model under Apache 2.0.

What the release provides

It uses a dense decoder-only Transformer with grouped-query attention and ReLU-squared activations. The base training used eight trillion tokens, followed by instruction tuning and reinforcement learning.AFM-4.5B official model card

How to use it well

Practical guidance. Compare task-specific adaptation with the unmodified instruction checkpoint before committing to training. Keep a clear separation between the source checkpoint, adaptation data and deployed revision. Review improvements on unseen examples as well as on the training domain.

Specifications & access

AFM-4.5B specifications
DetailValue and conditions
AccessDownloadable
Model identifiersarcee-ai/AFM-4.5B