This is a custom, optimized quantization of the Gemma 4 E4B (Abliterated) model. It was compiled specifically for execution on hardware with strict VRAM limitations, utilizing a hybrid importance matrix to prevent distribution collapse and preserve high-level reasoning.

The Importance Matrix (Imatrix) Architecture

To solve this, this model was quantized using a Hybrid Custom Imatrix. The calibration dataset was carefully blended using an specific ratio:

appx 93% High-Density Logic: Sourced from wikitext-2-raw to explicitly preserve reasoning, logic, etc.
appx 7% Custom Conversational Data: Sourced from highly specific, synthetic roleplay data.

The resulting imatrix.dat directed the llama-quantize tool to aggressively crush expendable weights while mathematically protecting the specific pathways needed for unfiltered reasoning.

Hardware & Performance Specifications

Format: GGUF (IQ3_M)
Size: ~4.4 GB
Target Hardware: 4GB VRAM GPUs.
Offloading Limits: Due to Gemma 4's architectural density, this model will exceed 4.0GB VRAM if fully offloaded. For optimal performance on 4GB cards, restrict GPU offloading to ~28 layers and utilize system RAM for the remainder.

Usage with Llama.cpp

To boot this model via the llama.cpp CLI with a 2048 context window and safe offload limits:

./llama-cli -m gemma-4-E4B-OBLITERATED-IMATRIX-IQ3_M.gguf -ngl 28 -c 2048 -cnv

levanell/gemma-4-OBLITERATED-IMATRIX-IQ3_M

Gemma 4 E4B - Obliterated & Imatrix Quantized (IQ3_M)

The Importance Matrix (Imatrix) Architecture

Hardware & Performance Specifications

Usage with Llama.cpp

levanell/
gemma-4-OBLITERATED-IMATRIX-IQ3_M