This is a layer-wise pruned variant of cerebras/Qwen3-Coder-REAP-25B-A3B resulting in a ~20B model with ~3B active parameters. It has not been fine-tuned yet.
Prune info:
Original model: 48 layers
New model: 38 Layers
result:
Similar model, it MUST be fine-tuned before usage as current performance is non-ideal. While it can absolutely be used as is, fine-tuning is needed.
