You can deploy this model and start earning money today!

johnjg75/
gemma-4-E4B-it-GGUF-pruned-Eng

Name: gemma-4-E4B-it-GGUF-pruned-Eng
Availability: InStock

gguf

google/gemma-4-E4B-it

Views

No views yet

🧠 Gemma-4 E4B (Pruned Embeddings - English Focus)

📌 Model Overview

This repository provides a pruned version of the Gemma 4 E4B GGUF model, optimized for significantly lower RAM usage during inference.

The original Gemma 4 architecture introduces Per-Layer Embeddings, which improves performance but dramatically increases memory requirements—especially for multi-language support.

To address this, this version removes approximately ~45% of embedding weights corresponding to unused languages, resulting in:

✅ Lower RAM usage
✅ Faster load times
✅ Smoother inference on constrained systems
⚠️ Reduced multilingual capability (English-focused)

⚙️ Key Features

🧩 Based on Gemma 4 E4B
✂️ Embedding pruning (~45% removed)
💾 Optimized for low-memory environments
⚡ GGUF format for compatibility with:
- llama.cpp
- koboldcpp
- text-generation-webui
🎯 Best suited for English-only or English-dominant tasks

🚨 Why This Exists

Gemma 4’s Per-Layer Embedding design causes a major increase in memory usage:

Each layer stores embedding-related data
Multilingual vocabularies multiply memory footprint
Results in RAM spikes during inference

💡 Solution

By removing embeddings for unused languages:

Memory footprint is drastically reduced
No impact on English performance (in typical use)
Enables running the model on hardware that would otherwise fail

johnjg75/gemma-4-E4B-it-GGUF-pruned-Eng