Downloads · 30 days
16
20% of all-time downloads
kzfastino/gemma-4-26B-A4B-it-pruned-106-experts
gemma-4-26B-A4B-it-pruned-106-experts is a image-text-to-text model from kzfastino. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as gemma.
One-shot whole-expert pruning of google/gemma-4-26B-A4B-it by vibepruner, produced live at the Google hackathon on 12 Sep 2026 on a single Lambda H100.
Downloads · 30 days
16
20% of all-time downloads
All-time downloads
80
Public
Parameters
21.9B
43.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors43.8 GB · 100%
From the Hugging Face model README
One-shot whole-expert pruning of google/gemma-4-26B-A4B-it by vibepruner,
produced live at the Google hackathon on 12 Sep 2026 on a single Lambda H100.
| original | this checkpoint | |
|---|---|---|
| parameters | 25.81B | 21.88B (−15.2%) |
| bf16 footprint | 51.6 GB | 43.8 GB |
| experts per layer | 128 | 106 in every layer (top-8 routing unchanged) |
| ARC-Challenge, 300 q | 95.3% | 94.0% (±1.3 pt 1σ) |
| top-1 agreement with original | — | 74.8% |
| KL(original ‖ pruned) | — | 0.96 nats/token |
Method. Router hooks record, per expert, the top-k gate weight and selection count over a calibration
set of wikitext prose plus the model's own chat-formatted answers. Saliency is
(assigned gate mass · ‖W_expert‖_F) / selections — a one-hook approximation of REAP's
g_j(x)·‖f_j(x)‖. Three rounds each removed the lowest-scoring 6.25% of remaining experts in every
layer via index_select on the stacked expert tensors and router rows, re-scoring between rounds. A fourth
round (−19.4%) dropped ARC-Challenge to 92.7% and was rejected.
Loading. Standard AutoModelForCausalLM.from_pretrained with transformers>=5.5; text_config.num_experts
is set to the reduced count.
Caveats. Perplexity is not reported: Gemma 4 instruct's next-token distribution has a hard floor at the logit softcap, which makes teacher-forced perplexity meaningless (≈2000 on wikitext for the unpruned model). No fine-tuning was applied. Per-token compute is unchanged (still top-8), so this is a footprint reduction, not a latency one.