OsaurusAI/gemma-4-26B-A4B-it-4bit

Gemma 4 26B-A4B-it — 4-bit (MLX)

Mixed-precision 4-bit quantization with verified vision tower weights

 

---

Model Details

Property	Value
Base Model	google/gemma-4-26B-A4B-it
Parameters	26B total, 4B active (Mixture of Experts)
Quantization	4-bit affine, mixed-precision (MLP layers kept at 8-bit)
Avg Bits/Weight	4.843
Model Size	14.8 GB
Architecture	Gemma 4 (text + vision)
Context Length	128K tokens
Vocabulary	262K tokens

Weight Verification

Every tensor in the vision tower was loaded and checked for max(abs(tensor)) > 0. Zero broken weights found.

Component	Tensor Count	Status
Vision Tower (SigLIP)	355	All non-zero
Language Model (MoE)	1,135	All non-zero
Total	1,490	All verified

Mixed-Precision Quantization

mlx-vlm's default quantization predicate automatically keeps MLP gate/up/down projections at 8-bit across all language model layers while quantizing attention and other weights to 4-bit. This improves quality over naive uniform 4-bit quantization.

Usage

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("OsaurusAI/gemma-4-26B-A4B-it-4bit")

# Text
prompt = apply_chat_template(processor, model.config, "Write a haiku about cats.")
output = generate(model, processor, prompt, max_tokens=200)
print(output.text)

# Vision
prompt = apply_chat_template(processor, model.config, "Describe this image.", num_images=1)
output = generate(model, processor, prompt, image="photo.jpg", max_tokens=200)
print(output.text)

Conversion

Converted from google/gemma-4-26B-A4B-it using mlx-vlm v0.4.4:

mlx_vlm.convert --hf-path google/gemma-4-26B-A4B-it \
  --mlx-path gemma-4-26b-a4b-it-4bit \
  -q --q-bits 4 --q-group-size 64 --q-mode affine --dtype bfloat16