Gemma 4 E4B Instruct Parity MLX (8bit)

This repository contains the canonical MLX side of the meshllm Gemma 4 parity pair.

  • Source checkpoint: google/gemma-4-E4B-it
  • Conversion path: original checkpoint -> MLX -> 8bit
  • Intended use: backend parity testing against the matching GGUF artifact

This artifact is not meant to be a generic Gemma 4 community release. It exists so that MLX and GGUF can be compared from the same original model lineage with minimal converter mismatch.

Canonical pair:

  • GGUF: meshllm/gemma-4-e4b-it-parity-q8_0-gguf
  • MLX: meshllm/gemma-4-e4b-it-parity-8bit-mlx

Latest trusted exact result:

Backend Model Exact
GGUF meshllm/gemma-4-e4b-it-parity-q8_0-gguf/gemma-4-e4b-it-q8_0.gguf PASS
MLX this repo PASS

Prompt comparison from local same-origin validation:

Prompt GGUF Q8_0 MLX 8bit
primary blue blue
alt-green green green
alt-red red red
capital-france Paris Paris
primary-colors red, green, blue red, green, blue
two-plus-two 4 4
largest-planet Jupiter Jupiter
breathing-gas Oxygen Oxygen
opposite-hot Cold Cold
banana-color Yellow Yellow
after-monday Tuesday Tuesday
Downloads last month
551
Safetensors
Model size
8B params
Tensor type
BF16
·
U32
·
MLX
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for meshllm/gemma-4-e4b-it-parity-8bit-mlx

Quantized
(73)
this model