license: apache-2.0
base_model: TrevorJS/gemma-4-E2B-it-uncensored
base_model_relation: quantized
library_name: gguf
tags:

• gguf
• llama.cpp
• gemma
• q4_0

gemma-4-E2B-it-uncensored — Q4_0 GGUF

Q4_0 quantization of TrevorJS/gemma-4-E2B-it-uncensored,
made for fast CPU inference on ARM phones (llama.cpp repacks Q4_0 weights for the NEON/DOTPROD kernels).

This is a modified version of the original work: the weights were re-quantized. Nothing else was changed.

• Source: gemma-4-E2B-it-uncensored-Q8_0.gguf from TrevorJS/gemma-4-E2B-it-uncensored-GGUF
• Tool: llama-quantize --allow-requantize … Q4_0 (llama.cpp b11228)
• Result: transformer layers in q4_0, token_embd and per_layer_token_embd in q6_K
• File: gemma-4-E2B-it-uncensored-Q4_0.gguf, 3,360,154,144 bytes
• SHA-256: 06a0d541e0aba58c8bfb5ea885493b75c291f176fb9ea154099cad7b80f87446

Measured on a Kirin 980 (4 threads, CPU only, llama.rn 0.12.9): prompt processing ~50 tok/s vs ~37 tok/s for the
author's Q4_K_M; generation ~10.6 tok/s vs ~9.5 tok/s (JSON-constrained).

License

Apache License 2.0, same as the original model. All credit for the model and the uncensoring method goes to
TrevorJS and, for the base model, Google (Gemma 4).