base_model: google/gemma-4-31B-it
library_name: mlx
pipeline_tag: text-generation
license: apache-2.0
tags:
- mlx
- jang
- jang-quantized
- JANG_2M
- mixed-precision
- apple-silicon
⚠️ Low-bit quality warning
This is an aggressive quantization (2-bit average). At this compression level, output quality degrades noticeably — responses may start coherent but degenerate into repetition or garbage tokens toward the end of longer generations. This is expected behavior for 2-bit quantization on this architecture.
Recommended for: experimentation, quick testing, extreme memory constraints. Not recommended for: production use, long-form generation, coding tasks.
For reliable output quality, use JANG_4M or higher profiles from this collection.
gemma-4-31B-it-JANG_2M
JANG adaptive mixed-precision MLX quantization produced via vmlx / jang-tools.
- Quantization: 3.75b avg, profile JANG_2M, method mse-all, calibration activations
- Profile: JANG_2M
- Format: JANG v2 MLX safetensors
- Compatible with: vmlx, MLX Studio, oMLX (with JANG patch)
Usage
vmlx (recommended)
pip install 'vmlx[jang]'
vmlx serve bearzi/gemma-4-31B-it-JANG_2M
Python
from jang_tools.loader import load_jang_model
from mlx_lm import generate
model, tokenizer = load_jang_model("bearzi/gemma-4-31B-it-JANG_2M")
messages = [{"role": "user", "content": "Hello"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=True))
About JANG
JANG (Jang Adaptive N-bit Grading) assigns different bit widths to different layer types — attention layers get more bits, MLP/expert layers compress harder. This preserves model coherence at aggressive compression levels where uniform quantization breaks down.
See JANG documentation and scores at jangq.ai.
Comparative benchmarks and feedback welcome — please open a discussion.