Gemma-4-26B-A4B REAP to Dense
A one-shot expert-pruned version of
google/gemma-4-26B-A4B-it,
produced with the REAP (Router-weighted Expert Activation Pruning)
pipeline as a ~4B dense conversion of the original 26B MoE.
Each layer keeps only the K = k = 8 most-salient experts exactly matching top_k_experts. So, every token is routed through all retained experts every time. The sparse routing degenerates into a dense FFN.
| Original | Pruned (this) | |
|---|---|---|
| Parameters (text MoE) | 25.2B | 4.39B |
| Experts per layer | 128 | 8 |
| Active per token | 8 | 8 |
| Weight size (bf16) | ~50 GB | 8.2 GB |
| Compression | — | 93.75% expert reduction |
The dense MLP block already present in each Gemma 4 layer is retained and its output is still summed with the MoE path, as in the original architecture.
- Base:
google/gemma-4-26B-A4B-it - Method: REAP (Router-weighted Expert Activation Pruning), layerwise error-accumulation calibration
Caveats
- It is intended as a research start point (e.g. via logit distillation)
- Vision tower and embeddings are unchanged from the original.
License
Apache 2.0, per the original Gemma 4 license. See license_link.