jang 1.4.0

JANG — Adaptive Mixed-Precision Quantization for Apple Silicon. The GGUF equivalent for MLX.

MLX Studio — the only app that natively supports JANG models

---

Compatibility Notice

JANG is a new quantization format. The following apps do NOT support JANG yet:

• LM Studio — does not support JANG
• Ollama — does not support JANG
• oMLX — does not support JANG
• Inferencer — does not support JANG

MLX Studio is currently the only app with native JANG support. You can also use the jang Python package directly (pip install "jang[mlx]").

Want JANG support in your favorite app? Ask the developers to add it! JANG is open-source (GitHub) and the format spec is public (FORMAT.md).

---

Jang Adaptive N-bit Grading

Mixed-Precision Quantization for Apple Silicon

The GGUF equivalent for MLX — models stay quantized in GPU memory at full Metal speed.

What is JANG?

JANG redistributes quantization bits based on tensor sensitivity. Critical layers (attention) get more bits, bulk layers (MLP) compensate — same total size, smarter allocation.

Like GGUF K-quants for MLX.

Results

4-bit: JANG_4K beats MLX 4-bit on MoE

Model	JANG_4K	MLX 4-bit	Size
Qwen3.5-122B MoE	94% MMLU	90%	69 GB vs 64 GB
Qwen3.5-35B MoE	84% MMLU	82%	16.7 GB vs 18 GB

2-bit: JANG doubles MLX on every model

Model	JANG_2S	MLX 2-bit	Size
Qwen3.5-122B MoE	84% MMLU	56%	38 GB vs 36 GB
Qwen3.5-35B MoE	62% MMLU	~20%	12 GB vs 10 GB
Qwen3.5-9B	36% MMLU	18%	3.5 GB vs 2.6 GB
Qwen3.5-4B	28% MMLU	14%	1.6 GB vs 1.3 GB

Install

pip install jang

For inference on Apple Silicon:

pip install "jang[mlx]"

Quick Start

Convert any model

# K-quant 4-bit (budget-neutral, same size as MLX, smarter)
jang convert Qwen/Qwen3.5-35B-A3B -p 4

# 2-bit for extreme compression
jang convert Qwen/Qwen3.5-122B-A10B -p 2

# Specific profile
jang convert model -p JANG_2S

Run inference

from jang_tools.loader import load_jang_model
from mlx_lm.sample_utils import make_sampler
from mlx_lm.generate import generate_step
import mlx.core as mx

model, tokenizer = load_jang_model("JANGQ-AI/Qwen3.5-122B-A10B-JANG_2S")
sampler = make_sampler(temp=0.7)

tokens = tokenizer.encode("What is photosynthesis?")
for tok, _ in generate_step(prompt=mx.array(tokens), model=model, max_tokens=200, sampler=sampler):
    t = tok.item() if hasattr(tok, 'item') else int(tok)
    print(tokenizer.decode([t]), end="", flush=True)
    if t == tokenizer.eos_token_id:
        break

Profiles

Profile	Type	Bits	Best for
JANG_4K	K-quant	4.0	Same size as MLX 4-bit, smarter
JANG_3K	K-quant	3.0	Same size as MLX 3-bit, smarter
JANG_2S	Profile	~2.1	Tightest 2-bit, near MLX 2-bit size
JANG_2M	Profile	~2.1	Balanced 2-bit
JANG_2L	Profile	~2.3	Quality 2-bit
JANG_1L	Profile	~2.2	Maximum quality 2-bit

Pre-quantized Models

Available on HuggingFace:

Model	Profile	MMLU	HumanEval	Size
Qwen3.5-122B-A10B	JANG_4K	94%	95%	69 GB
Qwen3.5-122B-A10B	JANG_2S	84%	90%	38 GB
Qwen3.5-35B-A3B	JANG_4K	84%	90%	16.7 GB
Qwen3.5-35B-A3B	JANG_2S	62%	—	12 GB

Supported Architectures

Dense Transformer, Mixture of Experts, Hybrid SSM, Linear Attention, MLA, Vision-Language, Mamba, FP8 source models.

Links

• GitHub | HuggingFace | MLX Studio | Format Spec

---

한국어

Apple Silicon을 위한 혼합정밀도 양자화

JANG이란?

JANG은 MLX를 위한 GGUF와 같은 역할을 하는 오픈소스 양자화 포맷입니다. 텐서 민감도에 따라 비트를 재분배합니다 — 중요 레이어(어텐션)에 더 많은 비트, 벌크 레이어(MLP)로 보상. 같은 크기, 더 스마트한 배분.

결과

4-bit: JANG_4K가 MLX 4-bit보다 우수 (MoE 모델)

모델	JANG_4K	MLX 4-bit	크기
Qwen3.5-122B MoE	94% MMLU	90%	69 GB vs 64 GB
Qwen3.5-35B MoE	84% MMLU	82%	16.7 GB vs 18 GB

2-bit: JANG이 모든 모델에서 MLX의 2배 성능

모델	JANG_2S	MLX 2-bit	크기
Qwen3.5-122B MoE	84% MMLU	56%	38 GB vs 36 GB
Qwen3.5-35B MoE	62% MMLU	~20%	12 GB vs 10 GB
Qwen3.5-9B	36% MMLU	18%	3.5 GB vs 2.6 GB
Qwen3.5-4B	28% MMLU	14%	1.6 GB vs 1.3 GB

설치

pip install jang                # 양자화 도구
pip install "jang[mlx]"         # + Apple Silicon 추론

모델 변환

# K-quant 4-bit (예산 중립, MLX와 같은 크기, 더 스마트)
jang convert Qwen/Qwen3.5-35B-A3B -p 4

# 극한 압축을 위한 2-bit
jang convert Qwen/Qwen3.5-122B-A10B -p 2

추론

from jang_tools.loader import load_jang_model
from mlx_lm.sample_utils import make_sampler
from mlx_lm.generate import generate_step
import mlx.core as mx

model, tokenizer = load_jang_model("JANGQ-AI/Qwen3.5-122B-A10B-JANG_2S")
sampler = make_sampler(temp=0.7)

tokens = tokenizer.encode("광합성이란 무엇인가요?")
for tok, _ in generate_step(prompt=mx.array(tokens), model=model, max_tokens=200, sampler=sampler):
    t = tok.item() if hasattr(tok, 'item') else int(tok)
    print(tokenizer.decode([t]), end="", flush=True)
    if t == tokenizer.eos_token_id:
        break

프로파일

프로파일	유형	비트	용도
JANG_4K	K-quant	4.0	MLX 4-bit과 같은 크기, 더 스마트
JANG_3K	K-quant	3.0	MLX 3-bit과 같은 크기, 더 스마트
JANG_2S	프로파일	~2.1	가장 타이트한 2-bit
JANG_1L	프로파일	~2.2	최대 품질 2-bit

사전 양자화 모델

HuggingFace에서 다운로드:

모델	프로파일	MMLU	HumanEval	크기
Qwen3.5-122B-A10B	JANG_4K	94%	95%	69 GB
Qwen3.5-122B-A10B	JANG_2S	84%	90%	38 GB
Qwen3.5-35B-A3B	JANG_4K	84%	90%	16.7 GB
Qwen3.5-35B-A3B	JANG_2S	62%	—	12 GB

호환성

현재 **MLX Studio**만 JANG 포맷을 기본 지원합니다. LM Studio, Ollama, oMLX, Inferencer 등은 아직 지원하지 않습니다. 좋아하는 앱의 개발자에게 JANG 지원을 요청해 주세요!

링크

• GitHub · HuggingFace · MLX Studio · PyPI

---

장진호 제작 · Created by Jinho Jang — jangq.ai