supergemma4-e4b-abliterated-mlx

SuperGemma4 E4B Abliterated MLX

This is the private Apple Silicon deployment build of supergemma4-e4b-abliterated, converted to MLX and quantized to a compact 4-bit format for fast local use on Mac Studio class hardware.

The original upstream checkpoint is google/gemma-4-E4B-it. This MLX package is the Apple Silicon deployment build of the final abliterated and tuned SuperGemma release derived from that Google E4B base.

If you want the strongest consumer-facing experience in this project line on Apple Silicon, this is the branch to pull first.

What You Get

• MLX-native 4-bit packaging
• compact single-file weight layout
• chat template preserved
• strong structured-output behavior inherited from the release candidate
• convenient path for local serving and Mac-based agent stacks

Derived From

• original upstream base: google/gemma-4-E4B-it
• source release: Jiunsong/supergemma4-e4b-abliterated

Release Highlights

The source release backing this MLX build achieved:

• release-quality score: 92.34
• exact-eval score: 98.50
• JSON exact-match: 100%
• tool-call accuracy: 90%
• exact code score: 100%
• exact bug-fix score: 100%
• long-context sanity: 100%

Serving and stability validation on the source candidate:

• direct reliability audit: 14/14
• repeat reliability probe: 90/90
• batched soak test: 12/12
• simple soak test: 6/6

Target Hardware

• Mac Studio
• Apple Silicon laptops and desktops
• MLX / vMLX local inference setups

Quick Start

from mlx_lm import load, generate

model, tokenizer = load("Jiunsong/supergemma4-e4b-abliterated-mlx")

messages = [
    {"role": "user", "content": "Write valid JSON with keys model and strength."}
]
prompt = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=False,
)

response = generate(model, tokenizer, prompt=prompt, max_tokens=128, verbose=False)
print(response)

Positioning

This branch is for users who want the SuperGemma4 E4B behavior in a lighter, Apple-friendly package that is easy to pull onto a Mac Studio for local testing and agent deployment.