license: apache-2.0
base_model:

• huihui-ai/Huihui-gemma-4-E2B-it-qat-q4_0-unquantized-abliterated
  tags:
• gguf
• gemma4
• qat
• q4_0
• abliterated
• uncensored
• llama.cpp

Huihui Gemma 4 E2B QAT — UD-Q4_K_XL-like

Source:

huihui-ai/Huihui-gemma-4-E2B-it-qat-q4_0-unquantized-abliterated

Files

• Huihui-gemma-4-E2B-it-qat-UD-Q4_K_XL-like.gguf — main model
• mmproj-BF16.gguf — BF16 multimodal projector
• imatrix_unsloth.gguf_file — exact Unsloth importance matrix used
• quantization_manifest.json — hashes and provenance

Conversion

• llama.cpp commit: fc6545d322a9ea6643f77439b31b66f403ba2cad
• Main quant type: Q4_0
• --pure: enabled
• imatrix: unsloth/gemma-4-E2B-it-GGUF/imatrix_unsloth.gguf_file
• mmproj: unsloth/gemma-4-E2B-it-qat-GGUF/mmproj-BF16.gguf

Important distinction

This is a publicly reproducible Unsloth-like QAT conversion, not the
official Unsloth UD-Q4_K_XL algorithm.

Unsloth's official Gemma 4 QAT GGUFs use additional scale/lattice recovery
logic when mapping the BF16 QAT lattice into llama.cpp's Q4_0 representation.
That exact recovery procedure is not exposed as a standard public
llama-quantize option.

llama.cpp example

llama-server \
  -m Huihui-gemma-4-E2B-it-qat-UD-Q4_K_XL-like.gguf \
  --mmproj mmproj-BF16.gguf \
  --jinja \
  -ngl 999